program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-milinternal", ""}, {"coremltools-version", "9.0"}})]
{
    func length_1<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(1)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_7")];
            tensor<fp16, [1, 1]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 1]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_6")];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 1]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 1]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_2_y_0 = const()[name = string("greater_equal_2_y_0"), val = int32(0)];
            tensor<bool, [1, 1]> greater_equal_2 = greater_equal(x = select_0, y = greater_equal_2_y_0)[name = string("greater_equal_2")];
            int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(32768)];
            tensor<int32, [1, 1]> add_2 = add(x = select_0, y = slice_by_index_2)[name = string("add_2")];
            tensor<int32, [1, 1]> select_2 = select(a = select_0, b = add_2, cond = greater_equal_2)[name = string("select_2")];
            int32 cos_1_cast_fp16_axis_1 = const()[name = string("cos_1_cast_fp16_axis_1"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_1, batch_dims = cos_1_batch_dims_0, indices = select_2, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_1 = const()[name = string("sin_1_cast_fp16_axis_1"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_1, batch_dims = sin_1_batch_dims_0, indices = select_2, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 1, 128, 1]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 1, 128, 1]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 1, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_5")];
            tensor<fp16, [1, 1, 1, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_63")];
            tensor<bool, [1, 1, 2048, 1]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_4")];
            tensor<fp16, [1, 2048, 1, 1]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_20_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_20 = read_state(input = key_cache)[name = string("coreml_update_state_20")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_21_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_21 = read_state(input = value_cache)[name = string("coreml_update_state_21")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_20)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_21)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_581_0, tensor<fp16, [1, 8, 128, 1]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_584_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_596_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_600, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_3_cast_fp16 = concat(axis = var_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_20)[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_22_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_22 = read_state(input = key_cache)[name = string("coreml_update_state_22")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_21)[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_23_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_23 = read_state(input = value_cache)[name = string("coreml_update_state_23")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_22)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_23)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_902_0, tensor<fp16, [1, 8, 128, 1]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_905_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_917_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_921, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_11_cast_fp16 = concat(axis = var_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_22)[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_24_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_24 = read_state(input = key_cache)[name = string("coreml_update_state_24")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_23)[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_25_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_25 = read_state(input = value_cache)[name = string("coreml_update_state_25")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_24)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_25)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1223_0, tensor<fp16, [1, 8, 128, 1]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1226_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1238_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_19_cast_fp16 = concat(axis = var_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_24)[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_26_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_26 = read_state(input = key_cache)[name = string("coreml_update_state_26")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_25)[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_27_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_27 = read_state(input = value_cache)[name = string("coreml_update_state_27")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_26)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_27)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1544_0, tensor<fp16, [1, 8, 128, 1]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1547_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1559_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_27_cast_fp16 = concat(axis = var_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_26)[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_28_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_28 = read_state(input = key_cache)[name = string("coreml_update_state_28")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_27)[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_29_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_29 = read_state(input = value_cache)[name = string("coreml_update_state_29")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_28)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_29)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1865_0, tensor<fp16, [1, 8, 128, 1]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1868_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1880_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_35_cast_fp16 = concat(axis = var_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_28)[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_30_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_30 = read_state(input = key_cache)[name = string("coreml_update_state_30")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_29)[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_31_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_31 = read_state(input = value_cache)[name = string("coreml_update_state_31")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_30)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_31)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2186_0, tensor<fp16, [1, 8, 128, 1]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2189_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_43_cast_fp16 = concat(axis = var_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_30)[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_32_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_32 = read_state(input = key_cache)[name = string("coreml_update_state_32")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_31)[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_33_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_33 = read_state(input = value_cache)[name = string("coreml_update_state_33")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_32)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_33)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2507_0, tensor<fp16, [1, 8, 128, 1]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2510_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2522_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_51_cast_fp16 = concat(axis = var_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_32)[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_34_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_34 = read_state(input = key_cache)[name = string("coreml_update_state_34")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_33)[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_35_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_35 = read_state(input = value_cache)[name = string("coreml_update_state_35")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_34)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_35)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2828_0, tensor<fp16, [1, 8, 128, 1]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2831_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2843_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_59_cast_fp16 = concat(axis = var_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_34)[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_36_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_36 = read_state(input = key_cache)[name = string("coreml_update_state_36")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_35)[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_37_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_37 = read_state(input = value_cache)[name = string("coreml_update_state_37")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_36)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_37)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3149_0, tensor<fp16, [1, 8, 128, 1]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3152_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3164_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_143_cast_fp16, y = var_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_67_cast_fp16 = concat(axis = var_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_36)[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_38_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_38 = read_state(input = key_cache)[name = string("coreml_update_state_38")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_37)[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_39_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_39 = read_state(input = value_cache)[name = string("coreml_update_state_39")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_38)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_39)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3470_0, tensor<fp16, [1, 8, 128, 1]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3473_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3485_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_75_cast_fp16 = concat(axis = var_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
    func length_128<ios18>(tensor<fp32, [1, 2048, 1, 128]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [128]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(128)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [128]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 128]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_27")];
            tensor<fp16, [1, 128]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 128]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_26")];
            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 128]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 128]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_12_y_0 = const()[name = string("greater_equal_12_y_0"), val = int32(0)];
            tensor<bool, [1, 128]> greater_equal_12 = greater_equal(x = select_0, y = greater_equal_12_y_0)[name = string("greater_equal_12")];
            int32 slice_by_index_12 = const()[name = string("slice_by_index_12"), val = int32(32768)];
            tensor<int32, [1, 128]> add_12 = add(x = select_0, y = slice_by_index_12)[name = string("add_12")];
            tensor<int32, [1, 128]> select_12 = select(a = select_0, b = add_12, cond = greater_equal_12)[name = string("select_12")];
            int32 cos_1_cast_fp16_axis_6 = const()[name = string("cos_1_cast_fp16_axis_6"), val = int32(0)];
            tensor<fp16, [1, 128, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_6, batch_dims = cos_1_batch_dims_0, indices = select_12, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_6 = const()[name = string("sin_1_cast_fp16_axis_6"), val = int32(0)];
            tensor<fp16, [1, 128, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_6, batch_dims = sin_1_batch_dims_0, indices = select_12, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 1, 128, 128]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 1, 128, 128]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 128, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 128, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_25")];
            tensor<fp16, [1, 1, 128, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 128]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_228")];
            tensor<bool, [1, 1, 2048, 128]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 128]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_24")];
            tensor<fp16, [1, 2048, 1, 128]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_120_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_120 = read_state(input = key_cache)[name = string("coreml_update_state_120")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_121_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_121 = read_state(input = value_cache)[name = string("coreml_update_state_121")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_120)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_121)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_581_0, tensor<fp16, [1, 8, 128, 128]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_584_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_596_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_15_cast_fp16 = softmax(axis = var_600, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_3_cast_fp16 = concat(axis = var_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_120)[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_122_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_122 = read_state(input = key_cache)[name = string("coreml_update_state_122")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_121)[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_123_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_123 = read_state(input = value_cache)[name = string("coreml_update_state_123")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_122)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_123)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_902_0, tensor<fp16, [1, 8, 128, 128]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_905_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_917_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_31_cast_fp16 = softmax(axis = var_921, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_11_cast_fp16 = concat(axis = var_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_122)[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_124_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_124 = read_state(input = key_cache)[name = string("coreml_update_state_124")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_123)[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_125_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_125 = read_state(input = value_cache)[name = string("coreml_update_state_125")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_124)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_125)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1223_0, tensor<fp16, [1, 8, 128, 128]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1226_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1238_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_19_cast_fp16 = concat(axis = var_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_124)[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_126_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_126 = read_state(input = key_cache)[name = string("coreml_update_state_126")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_125)[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_127_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_127 = read_state(input = value_cache)[name = string("coreml_update_state_127")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_126)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_127)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1544_0, tensor<fp16, [1, 8, 128, 128]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1547_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1559_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_27_cast_fp16 = concat(axis = var_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_126)[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_128_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_128 = read_state(input = key_cache)[name = string("coreml_update_state_128")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_127)[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_129_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_129 = read_state(input = value_cache)[name = string("coreml_update_state_129")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_128)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_129)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1865_0, tensor<fp16, [1, 8, 128, 128]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1868_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1880_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_35_cast_fp16 = concat(axis = var_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_128)[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_130_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_130 = read_state(input = key_cache)[name = string("coreml_update_state_130")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_129)[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_131_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_131 = read_state(input = value_cache)[name = string("coreml_update_state_131")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_130)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_131)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2186_0, tensor<fp16, [1, 8, 128, 128]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2189_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_43_cast_fp16 = concat(axis = var_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_130)[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_132_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_132 = read_state(input = key_cache)[name = string("coreml_update_state_132")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_131)[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_133_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_133 = read_state(input = value_cache)[name = string("coreml_update_state_133")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_132)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_133)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2507_0, tensor<fp16, [1, 8, 128, 128]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2510_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2522_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_51_cast_fp16 = concat(axis = var_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_132)[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_134_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_134 = read_state(input = key_cache)[name = string("coreml_update_state_134")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_133)[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_135_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_135 = read_state(input = value_cache)[name = string("coreml_update_state_135")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_134)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_135)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2828_0, tensor<fp16, [1, 8, 128, 128]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2831_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2843_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_59_cast_fp16 = concat(axis = var_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_134)[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_136_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_136 = read_state(input = key_cache)[name = string("coreml_update_state_136")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_135)[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_137_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_137 = read_state(input = value_cache)[name = string("coreml_update_state_137")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_136)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_137)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3149_0, tensor<fp16, [1, 8, 128, 128]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3152_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3164_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_143_cast_fp16, y = var_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_67_cast_fp16 = concat(axis = var_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_136)[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_138_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_138 = read_state(input = key_cache)[name = string("coreml_update_state_138")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_137)[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_139_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_139 = read_state(input = value_cache)[name = string("coreml_update_state_139")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_138)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_139)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3470_0, tensor<fp16, [1, 8, 128, 128]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3473_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3485_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_75_cast_fp16 = concat(axis = var_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
    func length_16<ios18>(tensor<fp32, [1, 2048, 1, 16]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [16]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(16)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [16]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 16]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_15")];
            tensor<fp16, [1, 16]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 16]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_14")];
            tensor<bool, [1, 16]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 16]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 16]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_6_y_0 = const()[name = string("greater_equal_6_y_0"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_6 = greater_equal(x = select_0, y = greater_equal_6_y_0)[name = string("greater_equal_6")];
            int32 slice_by_index_6 = const()[name = string("slice_by_index_6"), val = int32(32768)];
            tensor<int32, [1, 16]> add_6 = add(x = select_0, y = slice_by_index_6)[name = string("add_6")];
            tensor<int32, [1, 16]> select_6 = select(a = select_0, b = add_6, cond = greater_equal_6)[name = string("select_6")];
            int32 cos_1_cast_fp16_axis_3 = const()[name = string("cos_1_cast_fp16_axis_3"), val = int32(0)];
            tensor<fp16, [1, 16, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_3, batch_dims = cos_1_batch_dims_0, indices = select_6, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_3 = const()[name = string("sin_1_cast_fp16_axis_3"), val = int32(0)];
            tensor<fp16, [1, 16, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_3, batch_dims = sin_1_batch_dims_0, indices = select_6, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 1, 128, 16]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 1, 128, 16]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 16, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 16, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 16, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_13")];
            tensor<fp16, [1, 1, 16, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 16]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_129")];
            tensor<bool, [1, 1, 2048, 16]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 16]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_12")];
            tensor<fp16, [1, 2048, 1, 16]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_60)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_61)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_581_0, tensor<fp16, [1, 8, 128, 16]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_584_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_596_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_15_cast_fp16 = softmax(axis = var_600, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_3_cast_fp16 = concat(axis = var_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_60)[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_62_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_61)[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_63_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_62)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_63)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_902_0, tensor<fp16, [1, 8, 128, 16]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_905_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_917_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_31_cast_fp16 = softmax(axis = var_921, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_11_cast_fp16 = concat(axis = var_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_62)[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_64_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_64 = read_state(input = key_cache)[name = string("coreml_update_state_64")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_63)[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_65_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_65 = read_state(input = value_cache)[name = string("coreml_update_state_65")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_64)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_65)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1223_0, tensor<fp16, [1, 8, 128, 16]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1226_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1238_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_19_cast_fp16 = concat(axis = var_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_64)[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_66_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_66 = read_state(input = key_cache)[name = string("coreml_update_state_66")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_65)[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_67_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_67 = read_state(input = value_cache)[name = string("coreml_update_state_67")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_66)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_67)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1544_0, tensor<fp16, [1, 8, 128, 16]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1547_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1559_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_27_cast_fp16 = concat(axis = var_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_66)[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_68_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_68 = read_state(input = key_cache)[name = string("coreml_update_state_68")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_67)[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_69_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_69 = read_state(input = value_cache)[name = string("coreml_update_state_69")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_68)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_69)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1865_0, tensor<fp16, [1, 8, 128, 16]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1868_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1880_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_35_cast_fp16 = concat(axis = var_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_68)[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_70_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_70 = read_state(input = key_cache)[name = string("coreml_update_state_70")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_69)[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_71_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_71 = read_state(input = value_cache)[name = string("coreml_update_state_71")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_70)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_71)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2186_0, tensor<fp16, [1, 8, 128, 16]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2189_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_43_cast_fp16 = concat(axis = var_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_70)[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_72_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_72 = read_state(input = key_cache)[name = string("coreml_update_state_72")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_71)[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_73_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_73 = read_state(input = value_cache)[name = string("coreml_update_state_73")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_72)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_73)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2507_0, tensor<fp16, [1, 8, 128, 16]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2510_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2522_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_51_cast_fp16 = concat(axis = var_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_72)[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_74_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_74 = read_state(input = key_cache)[name = string("coreml_update_state_74")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_73)[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_75_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_75 = read_state(input = value_cache)[name = string("coreml_update_state_75")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_74)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_75)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2828_0, tensor<fp16, [1, 8, 128, 16]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2831_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2843_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_59_cast_fp16 = concat(axis = var_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_74)[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_76_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_76 = read_state(input = key_cache)[name = string("coreml_update_state_76")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_75)[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_77_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_77 = read_state(input = value_cache)[name = string("coreml_update_state_77")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_76)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_77)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3149_0, tensor<fp16, [1, 8, 128, 16]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3152_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3164_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_143_cast_fp16, y = var_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_67_cast_fp16 = concat(axis = var_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_76)[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_78_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_78 = read_state(input = key_cache)[name = string("coreml_update_state_78")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_77)[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_79_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_79 = read_state(input = value_cache)[name = string("coreml_update_state_79")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_78)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_79)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3470_0, tensor<fp16, [1, 8, 128, 16]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3473_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3485_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_75_cast_fp16 = concat(axis = var_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
    func length_32<ios18>(tensor<fp32, [1, 2048, 1, 32]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [32]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(32)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [32]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 32]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_19")];
            tensor<fp16, [1, 32]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 32]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_18")];
            tensor<bool, [1, 32]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 32]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 32]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_8_y_0 = const()[name = string("greater_equal_8_y_0"), val = int32(0)];
            tensor<bool, [1, 32]> greater_equal_8 = greater_equal(x = select_0, y = greater_equal_8_y_0)[name = string("greater_equal_8")];
            int32 slice_by_index_8 = const()[name = string("slice_by_index_8"), val = int32(32768)];
            tensor<int32, [1, 32]> add_8 = add(x = select_0, y = slice_by_index_8)[name = string("add_8")];
            tensor<int32, [1, 32]> select_8 = select(a = select_0, b = add_8, cond = greater_equal_8)[name = string("select_8")];
            int32 cos_1_cast_fp16_axis_4 = const()[name = string("cos_1_cast_fp16_axis_4"), val = int32(0)];
            tensor<fp16, [1, 32, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_4, batch_dims = cos_1_batch_dims_0, indices = select_8, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_4 = const()[name = string("sin_1_cast_fp16_axis_4"), val = int32(0)];
            tensor<fp16, [1, 32, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_4, batch_dims = sin_1_batch_dims_0, indices = select_8, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [1, 1, 128, 32]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 1, 128, 32]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 32, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 32, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 32, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_17")];
            tensor<fp16, [1, 1, 32, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 32]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_162")];
            tensor<bool, [1, 1, 2048, 32]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 32]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_16")];
            tensor<fp16, [1, 2048, 1, 32]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_80 = read_state(input = key_cache)[name = string("coreml_update_state_80")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_81 = read_state(input = value_cache)[name = string("coreml_update_state_81")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_80)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_81)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_581_0, tensor<fp16, [1, 8, 128, 32]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_584_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_596_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_15_cast_fp16 = softmax(axis = var_600, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_3_cast_fp16 = concat(axis = var_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_80)[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_82_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_82 = read_state(input = key_cache)[name = string("coreml_update_state_82")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_81)[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_83_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_83 = read_state(input = value_cache)[name = string("coreml_update_state_83")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_82)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_83)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_902_0, tensor<fp16, [1, 8, 128, 32]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_905_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_917_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_31_cast_fp16 = softmax(axis = var_921, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_11_cast_fp16 = concat(axis = var_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_82)[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_84_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_84 = read_state(input = key_cache)[name = string("coreml_update_state_84")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_83)[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_85_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_85 = read_state(input = value_cache)[name = string("coreml_update_state_85")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_84)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_85)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1223_0, tensor<fp16, [1, 8, 128, 32]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1226_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1238_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_19_cast_fp16 = concat(axis = var_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_84)[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_86_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_86 = read_state(input = key_cache)[name = string("coreml_update_state_86")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_85)[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_87_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_87 = read_state(input = value_cache)[name = string("coreml_update_state_87")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_86)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_87)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1544_0, tensor<fp16, [1, 8, 128, 32]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1547_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1559_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_27_cast_fp16 = concat(axis = var_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_86)[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_88_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_88 = read_state(input = key_cache)[name = string("coreml_update_state_88")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_87)[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_89_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_89 = read_state(input = value_cache)[name = string("coreml_update_state_89")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_88)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_89)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1865_0, tensor<fp16, [1, 8, 128, 32]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1868_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1880_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_35_cast_fp16 = concat(axis = var_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_88)[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_90_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_90 = read_state(input = key_cache)[name = string("coreml_update_state_90")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_89)[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_91_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_91 = read_state(input = value_cache)[name = string("coreml_update_state_91")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_90)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_91)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2186_0, tensor<fp16, [1, 8, 128, 32]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2189_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_43_cast_fp16 = concat(axis = var_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_90)[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_92_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_92 = read_state(input = key_cache)[name = string("coreml_update_state_92")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_91)[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_93_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_93 = read_state(input = value_cache)[name = string("coreml_update_state_93")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_92)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_93)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2507_0, tensor<fp16, [1, 8, 128, 32]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2510_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2522_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_51_cast_fp16 = concat(axis = var_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_92)[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_94_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_94 = read_state(input = key_cache)[name = string("coreml_update_state_94")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_93)[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_95_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_95 = read_state(input = value_cache)[name = string("coreml_update_state_95")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_94)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_95)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2828_0, tensor<fp16, [1, 8, 128, 32]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2831_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2843_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_59_cast_fp16 = concat(axis = var_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_94)[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_96_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_96 = read_state(input = key_cache)[name = string("coreml_update_state_96")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_95)[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_97_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_97 = read_state(input = value_cache)[name = string("coreml_update_state_97")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_96)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_97)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3149_0, tensor<fp16, [1, 8, 128, 32]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3152_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3164_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_143_cast_fp16, y = var_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_67_cast_fp16 = concat(axis = var_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_96)[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_98_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_98 = read_state(input = key_cache)[name = string("coreml_update_state_98")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_97)[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_99_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_99 = read_state(input = value_cache)[name = string("coreml_update_state_99")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_98)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_99)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3470_0, tensor<fp16, [1, 8, 128, 32]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3473_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3485_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_75_cast_fp16 = concat(axis = var_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
    func length_64<ios18>(tensor<fp32, [1, 2048, 1, 64]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [64]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(64)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [64]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 64]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_23")];
            tensor<fp16, [1, 64]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 64]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_22")];
            tensor<bool, [1, 64]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 64]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 64]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_10_y_0 = const()[name = string("greater_equal_10_y_0"), val = int32(0)];
            tensor<bool, [1, 64]> greater_equal_10 = greater_equal(x = select_0, y = greater_equal_10_y_0)[name = string("greater_equal_10")];
            int32 slice_by_index_10 = const()[name = string("slice_by_index_10"), val = int32(32768)];
            tensor<int32, [1, 64]> add_10 = add(x = select_0, y = slice_by_index_10)[name = string("add_10")];
            tensor<int32, [1, 64]> select_10 = select(a = select_0, b = add_10, cond = greater_equal_10)[name = string("select_10")];
            int32 cos_1_cast_fp16_axis_5 = const()[name = string("cos_1_cast_fp16_axis_5"), val = int32(0)];
            tensor<fp16, [1, 64, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_5, batch_dims = cos_1_batch_dims_0, indices = select_10, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_5 = const()[name = string("sin_1_cast_fp16_axis_5"), val = int32(0)];
            tensor<fp16, [1, 64, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_5, batch_dims = sin_1_batch_dims_0, indices = select_10, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 1, 128, 64]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [1, 1, 128, 64]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 64, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 64, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 64, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_21")];
            tensor<fp16, [1, 1, 64, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 64]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_195")];
            tensor<bool, [1, 1, 2048, 64]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 64]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_20")];
            tensor<fp16, [1, 2048, 1, 64]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_100 = read_state(input = key_cache)[name = string("coreml_update_state_100")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_101 = read_state(input = value_cache)[name = string("coreml_update_state_101")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_100)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_101)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_581_0, tensor<fp16, [1, 8, 128, 64]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_584_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_596_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_15_cast_fp16 = softmax(axis = var_600, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_3_cast_fp16 = concat(axis = var_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_100)[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_102_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_102 = read_state(input = key_cache)[name = string("coreml_update_state_102")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_101)[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_103_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_103 = read_state(input = value_cache)[name = string("coreml_update_state_103")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_102)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_103)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_902_0, tensor<fp16, [1, 8, 128, 64]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_905_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_917_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_31_cast_fp16 = softmax(axis = var_921, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_11_cast_fp16 = concat(axis = var_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_102)[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_104_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_104 = read_state(input = key_cache)[name = string("coreml_update_state_104")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_103)[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_105_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_105 = read_state(input = value_cache)[name = string("coreml_update_state_105")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_104)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_105)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1223_0, tensor<fp16, [1, 8, 128, 64]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1226_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1238_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_19_cast_fp16 = concat(axis = var_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_104)[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_106_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_106 = read_state(input = key_cache)[name = string("coreml_update_state_106")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_105)[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_107_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_107 = read_state(input = value_cache)[name = string("coreml_update_state_107")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_106)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_107)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1544_0, tensor<fp16, [1, 8, 128, 64]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1547_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1559_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_27_cast_fp16 = concat(axis = var_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_106)[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_108_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_108 = read_state(input = key_cache)[name = string("coreml_update_state_108")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_107)[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_109_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_109 = read_state(input = value_cache)[name = string("coreml_update_state_109")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_108)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_109)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1865_0, tensor<fp16, [1, 8, 128, 64]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1868_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1880_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_35_cast_fp16 = concat(axis = var_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_108)[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_110_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_110 = read_state(input = key_cache)[name = string("coreml_update_state_110")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_109)[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_111_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_111 = read_state(input = value_cache)[name = string("coreml_update_state_111")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_110)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_111)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2186_0, tensor<fp16, [1, 8, 128, 64]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2189_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_43_cast_fp16 = concat(axis = var_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_110)[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_112_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_112 = read_state(input = key_cache)[name = string("coreml_update_state_112")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_111)[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_113_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_113 = read_state(input = value_cache)[name = string("coreml_update_state_113")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_112)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_113)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2507_0, tensor<fp16, [1, 8, 128, 64]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2510_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2522_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_51_cast_fp16 = concat(axis = var_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_112)[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_114_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_114 = read_state(input = key_cache)[name = string("coreml_update_state_114")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_113)[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_115_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_115 = read_state(input = value_cache)[name = string("coreml_update_state_115")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_114)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_115)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2828_0, tensor<fp16, [1, 8, 128, 64]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2831_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2843_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_59_cast_fp16 = concat(axis = var_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_114)[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_116_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_116 = read_state(input = key_cache)[name = string("coreml_update_state_116")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_115)[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_117_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_117 = read_state(input = value_cache)[name = string("coreml_update_state_117")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_116)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_117)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3149_0, tensor<fp16, [1, 8, 128, 64]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3152_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3164_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_143_cast_fp16, y = var_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_67_cast_fp16 = concat(axis = var_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_116)[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_118_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_118 = read_state(input = key_cache)[name = string("coreml_update_state_118")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_117)[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_119_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_119 = read_state(input = value_cache)[name = string("coreml_update_state_119")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_118)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_119)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3470_0, tensor<fp16, [1, 8, 128, 64]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3473_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3485_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_75_cast_fp16 = concat(axis = var_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
    func length_8<ios18>(tensor<fp32, [1, 2048, 1, 8]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [8]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(8)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [8]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_11")];
            tensor<fp16, [1, 8]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 8]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_10")];
            tensor<bool, [1, 8]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 8]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 8]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_4_y_0 = const()[name = string("greater_equal_4_y_0"), val = int32(0)];
            tensor<bool, [1, 8]> greater_equal_4 = greater_equal(x = select_0, y = greater_equal_4_y_0)[name = string("greater_equal_4")];
            int32 slice_by_index_4 = const()[name = string("slice_by_index_4"), val = int32(32768)];
            tensor<int32, [1, 8]> add_4 = add(x = select_0, y = slice_by_index_4)[name = string("add_4")];
            tensor<int32, [1, 8]> select_4 = select(a = select_0, b = add_4, cond = greater_equal_4)[name = string("select_4")];
            int32 cos_1_cast_fp16_axis_2 = const()[name = string("cos_1_cast_fp16_axis_2"), val = int32(0)];
            tensor<fp16, [1, 8, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_2, batch_dims = cos_1_batch_dims_0, indices = select_4, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_2 = const()[name = string("sin_1_cast_fp16_axis_2"), val = int32(0)];
            tensor<fp16, [1, 8, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_2, batch_dims = sin_1_batch_dims_0, indices = select_4, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 1, 128, 8]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 1, 128, 8]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 8, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 8, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 8, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_9")];
            tensor<fp16, [1, 1, 8, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 8]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_96")];
            tensor<bool, [1, 1, 2048, 8]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 8]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_8")];
            tensor<fp16, [1, 2048, 1, 8]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_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_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_40_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_40 = read_state(input = key_cache)[name = string("coreml_update_state_40")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_41_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_41 = read_state(input = value_cache)[name = string("coreml_update_state_41")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_40)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_41)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_581_0, tensor<fp16, [1, 8, 128, 8]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            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_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_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_584_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_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_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_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_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_596_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_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_15_cast_fp16 = softmax(axis = var_600, 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_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), 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_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_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_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_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_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_40)[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_42_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_42 = read_state(input = key_cache)[name = string("coreml_update_state_42")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_41)[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_43_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_43 = read_state(input = value_cache)[name = string("coreml_update_state_43")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_42)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_43)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_902_0, tensor<fp16, [1, 8, 128, 8]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            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_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_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_905_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_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_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_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_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_917_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_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_31_cast_fp16 = softmax(axis = var_921, 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_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), 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_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_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_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_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_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_42)[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_44_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_44 = read_state(input = key_cache)[name = string("coreml_update_state_44")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_43)[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_45_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_45 = read_state(input = value_cache)[name = string("coreml_update_state_45")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_44)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_45)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1223_0, tensor<fp16, [1, 8, 128, 8]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            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_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_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_1226_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_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_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_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_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_1238_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_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, 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_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), 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_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_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_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_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_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_44)[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_46_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_46 = read_state(input = key_cache)[name = string("coreml_update_state_46")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_45)[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_47_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_47 = read_state(input = value_cache)[name = string("coreml_update_state_47")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_46)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_47)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1544_0, tensor<fp16, [1, 8, 128, 8]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            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_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_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_1547_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_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_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_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_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_1559_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_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, 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_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), 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_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_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_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_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_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_46)[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_48_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_48 = read_state(input = key_cache)[name = string("coreml_update_state_48")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_47)[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_49_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_49 = read_state(input = value_cache)[name = string("coreml_update_state_49")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_48)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_49)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1865_0, tensor<fp16, [1, 8, 128, 8]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            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_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_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_1868_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_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_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_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_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_1880_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_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, 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_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), 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_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_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_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_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_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_48)[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_50_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_50 = read_state(input = key_cache)[name = string("coreml_update_state_50")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_49)[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_51_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_51 = read_state(input = value_cache)[name = string("coreml_update_state_51")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_50)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_51)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2186_0, tensor<fp16, [1, 8, 128, 8]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            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_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_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_2189_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_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_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_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_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_2201_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_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, 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_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), 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_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_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_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_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_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_50)[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_52_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_52 = read_state(input = key_cache)[name = string("coreml_update_state_52")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_51)[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_53_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_53 = read_state(input = value_cache)[name = string("coreml_update_state_53")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_52)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_53)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2507_0, tensor<fp16, [1, 8, 128, 8]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            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_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_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_2510_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_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_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_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_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_2522_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_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, 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_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), 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_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_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_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_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_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_52)[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_54_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_54 = read_state(input = key_cache)[name = string("coreml_update_state_54")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_53)[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_55_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_55 = read_state(input = value_cache)[name = string("coreml_update_state_55")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_54)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_55)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2828_0, tensor<fp16, [1, 8, 128, 8]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            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_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_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_2831_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_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_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_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_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_2843_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_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, 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_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), 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_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_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_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_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_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_54)[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_56_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_55)[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_57_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_56)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_57)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3149_0, tensor<fp16, [1, 8, 128, 8]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            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_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_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_3152_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_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_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_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_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_3164_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_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, 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_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), 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_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_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_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_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_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_56)[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_58_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_57)[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_59_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_58)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_59)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3470_0, tensor<fp16, [1, 8, 128, 8]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            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_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_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_3473_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_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_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_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_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_3485_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_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_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_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), 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_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
    func main<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [10, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [10, 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")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(1)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_308 = const()[name = string("op_308"), val = int32(0)];
            bool var_310_exclusive_0 = const()[name = string("op_310_exclusive_0"), val = bool(false)];
            bool var_310_reverse_0 = const()[name = string("op_310_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_310_cast_fp16 = cumsum(axis = var_308, exclusive = var_310_exclusive_0, reverse = var_310_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_310_cast_fp16")];
            fp16 var_312_promoted_to_fp16 = const()[name = string("op_312_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_310_cast_fp16, y = var_312_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = position_offsets_cast_fp16)[name = string("op_315_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_3")];
            tensor<fp16, [1, 1]> position_ids_1_cast_fp16 = add(x = var_315_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 1]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_2")];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 1]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 1]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260885696)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 1]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(32768)];
            tensor<int32, [1, 1]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
            tensor<int32, [1, 1]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 cos_1_cast_fp16_axis_0 = const()[name = string("cos_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_0, batch_dims = cos_1_batch_dims_0, indices = select_0_1, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269274368)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_0 = const()[name = string("sin_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_0, batch_dims = sin_1_batch_dims_0, indices = select_0_1, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_334_perm_0 = const()[name = string("op_334_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_336_axes_0 = const()[name = string("op_336_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_334_cast_fp16 = transpose(perm = var_334_perm_0, x = cos_1_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 1, 128, 1]> var_336_cast_fp16 = expand_dims(axes = var_336_axes_0, x = var_334_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<int32, [3]> var_341_perm_0 = const()[name = string("op_341_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_343_axes_0 = const()[name = string("op_343_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_341_cast_fp16 = transpose(perm = var_341_perm_0, x = sin_1_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 1, 128, 1]> var_343_cast_fp16 = expand_dims(axes = var_343_axes_0, x = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_362_axes_0 = const()[name = string("op_362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_362 = expand_dims(axes = var_362_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_362")];
            tensor<int32, [1, 1, 2048]> var_355 = const()[name = string("op_355"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277663040)))];
            tensor<bool, [1, 1, 2048]> var_363 = greater(x = var_355, y = var_362)[name = string("op_363")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_363_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_363)[name = string("cast_1")];
            tensor<fp16, [1, 1, 1, 2048]> var_370_cast_fp16 = expand_dims(axes = var_370_axes_0, x = var_363_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_374_promoted_to_fp16 = const()[name = string("op_374_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_370_cast_fp16)[name = string("transpose_30")];
            tensor<bool, [1, 1, 2048, 1]> var_375_cast_fp16 = equal(x = mask_cast_fp16, y = var_374_promoted_to_fp16)[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_376_to_fp16, cond = var_375_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_0")];
            tensor<fp16, [1, 2048, 1, 1]> var_386_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_386_cast_fp16")];
            int32 var_384 = const()[name = string("op_384"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_1_cast_fp16 = concat(axis = var_384, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_386_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(277671296)))];
            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_396_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_407_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_407_cast_fp16_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = out_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277679552)))];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286068224)))];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287116864)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_407_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_464_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_471_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_475_cast_fp16 = mul(x = x_1_cast_fp16, y = var_336_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [2]> var_476_split_sizes_0 = const()[name = string("op_476_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_476_axis_0 = const()[name = string("op_476_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_476_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_476_cast_fp16_1 = split(axis = var_476_axis_0, split_sizes = var_476_split_sizes_0, x = x_1_cast_fp16)[name = string("op_476_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_478_cast_fp16 = mul(x = var_476_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_478_cast_fp16")];
            int32 var_480 = const()[name = string("op_480"), val = int32(-2)];
            bool var_481_interleave_0 = const()[name = string("op_481_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_481_cast_fp16 = concat(axis = var_480, interleave = var_481_interleave_0, values = (var_478_cast_fp16, var_476_cast_fp16_0))[name = string("op_481_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_482_cast_fp16 = mul(x = var_481_cast_fp16, y = var_343_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_475_cast_fp16, y = var_482_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_488_cast_fp16 = mul(x = var_464_cast_fp16, y = var_336_cast_fp16)[name = string("op_488_cast_fp16")];
            tensor<int32, [2]> var_489_split_sizes_0 = const()[name = string("op_489_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_489_axis_0 = const()[name = string("op_489_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_489_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_489_cast_fp16_1 = split(axis = var_489_axis_0, split_sizes = var_489_split_sizes_0, x = var_464_cast_fp16)[name = string("op_489_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_491_cast_fp16 = mul(x = var_489_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            int32 var_493 = const()[name = string("op_493"), val = int32(-2)];
            bool var_494_interleave_0 = const()[name = string("op_494_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_494_cast_fp16 = concat(axis = var_493, interleave = var_494_interleave_0, values = (var_491_cast_fp16, var_489_cast_fp16_0))[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_495_cast_fp16 = mul(x = var_494_cast_fp16, y = var_343_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_488_cast_fp16, y = var_495_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_0_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_0 = read_state(input = key_cache)[name = string("coreml_update_state_0")];
            tensor<fp16, [10, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_471_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_1_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_1 = read_state(input = value_cache)[name = string("coreml_update_state_1")];
            tensor<int32, [4]> var_565_begin_0 = const()[name = string("op_565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_565_end_0 = const()[name = string("op_565_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_565_end_mask_0 = const()[name = string("op_565_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, x = coreml_update_state_0)[name = string("op_565_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_axis_0 = const()[name = string("op_568_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = tile_0, x = var_565_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<int32, [4]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_575_end_0 = const()[name = string("op_575_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = var_575_end_0, end_mask = var_575_end_mask_0, x = coreml_update_state_1)[name = string("op_575_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_578_axis_0 = const()[name = string("op_578_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_578_cast_fp16_1 = split(axis = var_578_axis_0, split_sizes = tile_1, x = var_575_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<int32, [2]> var_581_split_sizes_0 = const()[name = string("op_581_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_581_axis_0 = const()[name = string("op_581_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_581_0, tensor<fp16, [1, 8, 128, 1]> var_581_1 = split(axis = var_581_axis_0, split_sizes = var_581_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_581")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_568_cast_fp16_0, y = var_581_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_584_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_588 = const()[name = string("op_588"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_588, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(true)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_578_cast_fp16_0)[name = string("op_594_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_568_cast_fp16_1, y = var_581_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_596_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_600 = const()[name = string("op_600"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_600, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_578_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_608 = const()[name = string("op_608"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_3_cast_fp16 = concat(axis = var_608, interleave = attn_output_3_interleave_0, values = (var_594_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_612_perm_0 = const()[name = string("op_612_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_612_cast_fp16 = transpose(perm = var_612_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_612_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288165504)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_645_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            int32 var_643 = const()[name = string("op_643"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_5_cast_fp16 = concat(axis = var_643, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_645_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(296554176)))];
            fp16 var_655_to_fp16 = const()[name = string("op_655_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_655_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_666_split_sizes_0 = const()[name = string("op_666_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_666_axis_0 = const()[name = string("op_666_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_666_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_666_cast_fp16_1 = split(axis = var_666_axis_0, split_sizes = var_666_split_sizes_0, x = out_3_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296562432)))];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_683_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321728320)))];
            tensor<int32, [2]> var_689_strides_0 = const()[name = string("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_689_pad_type_0 = const()[name = string("op_689_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = string("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = string("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_689_groups_0 = const()[name = string("op_689_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_689_cast_fp16 = conv(dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_666_cast_fp16_0)[name = string("op_689_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_683_cast_fp16, y = var_689_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346894208)))];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_707_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_707_cast_fp16")];
            int32 var_705 = const()[name = string("op_705"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_9_cast_fp16 = concat(axis = var_705, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_707_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(372060096)))];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_717_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_728_split_sizes_0 = const()[name = string("op_728_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_728_axis_0 = const()[name = string("op_728_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_728_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_728_cast_fp16_1 = split(axis = var_728_axis_0, split_sizes = var_728_split_sizes_0, x = out_5_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372068352)))];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380457024)))];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = var_728_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_cast_fp16, x = var_728_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_785_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_792_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_796_cast_fp16 = mul(x = x_11_cast_fp16, y = var_336_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<int32, [2]> var_797_split_sizes_0 = const()[name = string("op_797_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_797_axis_0 = const()[name = string("op_797_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_797_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_797_cast_fp16_1 = split(axis = var_797_axis_0, split_sizes = var_797_split_sizes_0, x = x_11_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_799_cast_fp16 = mul(x = var_797_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_799_cast_fp16")];
            int32 var_801 = const()[name = string("op_801"), val = int32(-2)];
            bool var_802_interleave_0 = const()[name = string("op_802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_802_cast_fp16 = concat(axis = var_801, interleave = var_802_interleave_0, values = (var_799_cast_fp16, var_797_cast_fp16_0))[name = string("op_802_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_803_cast_fp16 = mul(x = var_802_cast_fp16, y = var_343_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_796_cast_fp16, y = var_803_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_809_cast_fp16 = mul(x = var_785_cast_fp16, y = var_336_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [2]> var_810_split_sizes_0 = const()[name = string("op_810_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_810_axis_0 = const()[name = string("op_810_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_810_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_810_cast_fp16_1 = split(axis = var_810_axis_0, split_sizes = var_810_split_sizes_0, x = var_785_cast_fp16)[name = string("op_810_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_812_cast_fp16 = mul(x = var_810_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_812_cast_fp16")];
            int32 var_814 = const()[name = string("op_814"), val = int32(-2)];
            bool var_815_interleave_0 = const()[name = string("op_815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_815_cast_fp16 = concat(axis = var_814, interleave = var_815_interleave_0, values = (var_812_cast_fp16, var_810_cast_fp16_0))[name = string("op_815_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_816_cast_fp16 = mul(x = var_815_cast_fp16, y = var_343_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_809_cast_fp16, y = var_816_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_0)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_2_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_2 = read_state(input = key_cache)[name = string("coreml_update_state_2")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_792_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_1)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_3_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_3 = read_state(input = value_cache)[name = string("coreml_update_state_3")];
            tensor<int32, [4]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = coreml_update_state_2)[name = string("op_886_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_889_axis_0 = const()[name = string("op_889_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_889_cast_fp16_1 = split(axis = var_889_axis_0, split_sizes = tile_2, x = var_886_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<int32, [4]> var_896_begin_0 = const()[name = string("op_896_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_896_end_0 = const()[name = string("op_896_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_896_end_mask_0 = const()[name = string("op_896_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_896_cast_fp16 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = coreml_update_state_3)[name = string("op_896_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_899_axis_0 = const()[name = string("op_899_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_899_cast_fp16_1 = split(axis = var_899_axis_0, split_sizes = tile_3, x = var_896_cast_fp16)[name = string("op_899_cast_fp16")];
            tensor<int32, [2]> var_902_split_sizes_0 = const()[name = string("op_902_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_902_axis_0 = const()[name = string("op_902_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_902_0, tensor<fp16, [1, 8, 128, 1]> var_902_1 = split(axis = var_902_axis_0, split_sizes = var_902_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_902")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_889_cast_fp16_0, y = var_902_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_905_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_909 = const()[name = string("op_909"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_909, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_915_transpose_x_1 = const()[name = string("op_915_transpose_x_1"), val = bool(true)];
            bool var_915_transpose_y_1 = const()[name = string("op_915_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_915_cast_fp16 = matmul(transpose_x = var_915_transpose_x_1, transpose_y = var_915_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_899_cast_fp16_0)[name = string("op_915_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_889_cast_fp16_1, y = var_902_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_917_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_921 = const()[name = string("op_921"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_921, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_899_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_929 = const()[name = string("op_929"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_11_cast_fp16 = concat(axis = var_929, interleave = attn_output_11_interleave_0, values = (var_915_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_933_cast_fp16 = transpose(perm = var_933_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_933_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381505664)))];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_966_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_966_cast_fp16")];
            int32 var_964 = const()[name = string("op_964"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_13_cast_fp16 = concat(axis = var_964, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_966_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(389894336)))];
            fp16 var_976_to_fp16 = const()[name = string("op_976_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_976_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_987_split_sizes_0 = const()[name = string("op_987_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_987_axis_0 = const()[name = string("op_987_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_987_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_987_cast_fp16_1 = split(axis = var_987_axis_0, split_sizes = var_987_split_sizes_0, x = out_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389902592)))];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = var_987_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1004_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1004_cast_fp16")];
            tensor<int32, [2]> var_1010_strides_0 = const()[name = string("op_1010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1010_pad_type_0 = const()[name = string("op_1010_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1010_pad_0 = const()[name = string("op_1010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1010_dilations_0 = const()[name = string("op_1010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1010_groups_0 = const()[name = string("op_1010_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1010_cast_fp16 = conv(dilations = var_1010_dilations_0, groups = var_1010_groups_0, pad = var_1010_pad_0, pad_type = var_1010_pad_type_0, strides = var_1010_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_987_cast_fp16_0)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_1010_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415068480)))];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1028_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1028_cast_fp16")];
            int32 var_1026 = const()[name = string("op_1026"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_17_cast_fp16 = concat(axis = var_1026, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1028_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(440234368)))];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1038_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1049_split_sizes_0 = const()[name = string("op_1049_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1049_axis_0 = const()[name = string("op_1049_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1049_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1049_cast_fp16_1 = split(axis = var_1049_axis_0, split_sizes = var_1049_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440242624)))];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448631296)))];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = var_1049_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_cast_fp16, x = var_1049_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1106_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1113_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1113_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1117_cast_fp16 = mul(x = x_21_cast_fp16, y = var_336_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [2]> var_1118_split_sizes_0 = const()[name = string("op_1118_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1118_axis_0 = const()[name = string("op_1118_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1118_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1118_cast_fp16_1 = split(axis = var_1118_axis_0, split_sizes = var_1118_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1120_cast_fp16 = mul(x = var_1118_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1120_cast_fp16")];
            int32 var_1122 = const()[name = string("op_1122"), val = int32(-2)];
            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1123_cast_fp16 = concat(axis = var_1122, interleave = var_1123_interleave_0, values = (var_1120_cast_fp16, var_1118_cast_fp16_0))[name = string("op_1123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1124_cast_fp16 = mul(x = var_1123_cast_fp16, y = var_343_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1124_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1130_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_336_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [2]> var_1131_split_sizes_0 = const()[name = string("op_1131_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1131_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1131_cast_fp16_1 = split(axis = var_1131_axis_0, split_sizes = var_1131_split_sizes_0, x = var_1106_cast_fp16)[name = string("op_1131_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1133_cast_fp16")];
            int32 var_1135 = const()[name = string("op_1135"), val = int32(-2)];
            bool var_1136_interleave_0 = const()[name = string("op_1136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1136_cast_fp16 = concat(axis = var_1135, interleave = var_1136_interleave_0, values = (var_1133_cast_fp16, var_1131_cast_fp16_0))[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1137_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_343_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1130_cast_fp16, y = var_1137_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_2)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_4_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_4 = read_state(input = key_cache)[name = string("coreml_update_state_4")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1113_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_3)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_5_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_5 = read_state(input = value_cache)[name = string("coreml_update_state_5")];
            tensor<int32, [4]> var_1207_begin_0 = const()[name = string("op_1207_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1207_end_0 = const()[name = string("op_1207_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1207_end_mask_0 = const()[name = string("op_1207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1207_cast_fp16 = slice_by_index(begin = var_1207_begin_0, end = var_1207_end_0, end_mask = var_1207_end_mask_0, x = coreml_update_state_4)[name = string("op_1207_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1210_axis_0 = const()[name = string("op_1210_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1210_cast_fp16_1 = split(axis = var_1210_axis_0, split_sizes = tile_4, x = var_1207_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<int32, [4]> var_1217_begin_0 = const()[name = string("op_1217_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1217_end_0 = const()[name = string("op_1217_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1217_end_mask_0 = const()[name = string("op_1217_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1217_cast_fp16 = slice_by_index(begin = var_1217_begin_0, end = var_1217_end_0, end_mask = var_1217_end_mask_0, x = coreml_update_state_5)[name = string("op_1217_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1220_axis_0 = const()[name = string("op_1220_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1220_cast_fp16_1 = split(axis = var_1220_axis_0, split_sizes = tile_5, x = var_1217_cast_fp16)[name = string("op_1220_cast_fp16")];
            tensor<int32, [2]> var_1223_split_sizes_0 = const()[name = string("op_1223_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1223_axis_0 = const()[name = string("op_1223_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1223_0, tensor<fp16, [1, 8, 128, 1]> var_1223_1 = split(axis = var_1223_axis_0, split_sizes = var_1223_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1223")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1210_cast_fp16_0, y = var_1223_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1226_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1230 = const()[name = string("op_1230"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1230, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1236_transpose_x_1 = const()[name = string("op_1236_transpose_x_1"), val = bool(true)];
            bool var_1236_transpose_y_1 = const()[name = string("op_1236_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1236_cast_fp16 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1220_cast_fp16_0)[name = string("op_1236_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1210_cast_fp16_1, y = var_1223_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1238_to_fp16 = const()[name = string("op_1238_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1238_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1242 = const()[name = string("op_1242"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1242, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1220_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1250 = const()[name = string("op_1250"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_19_cast_fp16 = concat(axis = var_1250, interleave = attn_output_19_interleave_0, values = (var_1236_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1254_perm_0 = const()[name = string("op_1254_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1254_cast_fp16 = transpose(perm = var_1254_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1254_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449679936)))];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1287_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1287_cast_fp16")];
            int32 var_1285 = const()[name = string("op_1285"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_21_cast_fp16 = concat(axis = var_1285, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1287_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(458068608)))];
            fp16 var_1297_to_fp16 = const()[name = string("op_1297_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1297_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1308_split_sizes_0 = const()[name = string("op_1308_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1308_axis_0 = const()[name = string("op_1308_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1308_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1308_cast_fp16_1 = split(axis = var_1308_axis_0, split_sizes = var_1308_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458076864)))];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = var_1308_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1325_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [2]> var_1331_strides_0 = const()[name = string("op_1331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1331_pad_type_0 = const()[name = string("op_1331_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1331_pad_0 = const()[name = string("op_1331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1331_dilations_0 = const()[name = string("op_1331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1331_groups_0 = const()[name = string("op_1331_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1331_cast_fp16 = conv(dilations = var_1331_dilations_0, groups = var_1331_groups_0, pad = var_1331_pad_0, pad_type = var_1331_pad_type_0, strides = var_1331_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1308_cast_fp16_0)[name = string("op_1331_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_1331_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483242752)))];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1349_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1349_cast_fp16")];
            int32 var_1347 = const()[name = string("op_1347"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_25_cast_fp16 = concat(axis = var_1347, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1349_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(508408640)))];
            fp16 var_1359_to_fp16 = const()[name = string("op_1359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1359_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1370_split_sizes_0 = const()[name = string("op_1370_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1370_axis_0 = const()[name = string("op_1370_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1370_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1370_cast_fp16_1 = split(axis = var_1370_axis_0, split_sizes = var_1370_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508416896)))];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516805568)))];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = var_1370_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_cast_fp16, x = var_1370_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1427_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1434_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1438_cast_fp16 = mul(x = x_31_cast_fp16, y = var_336_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [2]> var_1439_split_sizes_0 = const()[name = string("op_1439_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1439_axis_0 = const()[name = string("op_1439_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1439_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1439_cast_fp16_1 = split(axis = var_1439_axis_0, split_sizes = var_1439_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1439_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1441_cast_fp16 = mul(x = var_1439_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
            int32 var_1443 = const()[name = string("op_1443"), val = int32(-2)];
            bool var_1444_interleave_0 = const()[name = string("op_1444_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1444_cast_fp16 = concat(axis = var_1443, interleave = var_1444_interleave_0, values = (var_1441_cast_fp16, var_1439_cast_fp16_0))[name = string("op_1444_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1445_cast_fp16 = mul(x = var_1444_cast_fp16, y = var_343_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1438_cast_fp16, y = var_1445_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1451_cast_fp16 = mul(x = var_1427_cast_fp16, y = var_336_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [2]> var_1452_split_sizes_0 = const()[name = string("op_1452_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1452_axis_0 = const()[name = string("op_1452_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1452_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1452_cast_fp16_1 = split(axis = var_1452_axis_0, split_sizes = var_1452_split_sizes_0, x = var_1427_cast_fp16)[name = string("op_1452_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1454_cast_fp16 = mul(x = var_1452_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1454_cast_fp16")];
            int32 var_1456 = const()[name = string("op_1456"), val = int32(-2)];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1457_cast_fp16 = concat(axis = var_1456, interleave = var_1457_interleave_0, values = (var_1454_cast_fp16, var_1452_cast_fp16_0))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = var_343_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1451_cast_fp16, y = var_1458_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_4)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_6_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_6 = read_state(input = key_cache)[name = string("coreml_update_state_6")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1434_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_5)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_7_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_7 = read_state(input = value_cache)[name = string("coreml_update_state_7")];
            tensor<int32, [4]> var_1528_begin_0 = const()[name = string("op_1528_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1528_end_0 = const()[name = string("op_1528_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1528_end_mask_0 = const()[name = string("op_1528_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = coreml_update_state_6)[name = string("op_1528_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1531_axis_0 = const()[name = string("op_1531_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1531_cast_fp16_1 = split(axis = var_1531_axis_0, split_sizes = tile_6, x = var_1528_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [4]> var_1538_begin_0 = const()[name = string("op_1538_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1538_end_0 = const()[name = string("op_1538_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1538_end_mask_0 = const()[name = string("op_1538_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1538_cast_fp16 = slice_by_index(begin = var_1538_begin_0, end = var_1538_end_0, end_mask = var_1538_end_mask_0, x = coreml_update_state_7)[name = string("op_1538_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1541_axis_0 = const()[name = string("op_1541_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1541_cast_fp16_1 = split(axis = var_1541_axis_0, split_sizes = tile_7, x = var_1538_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1544_0, tensor<fp16, [1, 8, 128, 1]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1544")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1531_cast_fp16_0, y = var_1544_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1547_to_fp16 = const()[name = string("op_1547_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1547_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1551 = const()[name = string("op_1551"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1551, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1557_transpose_x_1 = const()[name = string("op_1557_transpose_x_1"), val = bool(true)];
            bool var_1557_transpose_y_1 = const()[name = string("op_1557_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1557_cast_fp16 = matmul(transpose_x = var_1557_transpose_x_1, transpose_y = var_1557_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1541_cast_fp16_0)[name = string("op_1557_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1531_cast_fp16_1, y = var_1544_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1559_to_fp16 = const()[name = string("op_1559_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1559_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1563 = const()[name = string("op_1563"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1563, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1541_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1571 = const()[name = string("op_1571"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_27_cast_fp16 = concat(axis = var_1571, interleave = attn_output_27_interleave_0, values = (var_1557_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1575_perm_0 = const()[name = string("op_1575_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1575_cast_fp16 = transpose(perm = var_1575_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1575_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_cast_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1608_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1608_cast_fp16")];
            int32 var_1606 = const()[name = string("op_1606"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_29_cast_fp16 = concat(axis = var_1606, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1608_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(517854208)))];
            fp16 var_1618_to_fp16 = const()[name = string("op_1618_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1618_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1629_split_sizes_0 = const()[name = string("op_1629_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1629_axis_0 = const()[name = string("op_1629_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1629_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1629_cast_fp16_1 = split(axis = var_1629_axis_0, split_sizes = var_1629_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517862464)))];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1646_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543028352)))];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1629_cast_fp16_0)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_1652_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1670_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1670_cast_fp16")];
            int32 var_1668 = const()[name = string("op_1668"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_33_cast_fp16 = concat(axis = var_1668, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1670_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(568194240)))];
            fp16 var_1680_to_fp16 = const()[name = string("op_1680_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1680_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1691_split_sizes_0 = const()[name = string("op_1691_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1691_axis_0 = const()[name = string("op_1691_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1691_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1691_cast_fp16_1 = split(axis = var_1691_axis_0, split_sizes = var_1691_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568202496)))];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576591168)))];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = var_1691_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_cast_fp16, x = var_1691_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1748_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1755_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1759_cast_fp16 = mul(x = x_41_cast_fp16, y = var_336_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [2]> var_1760_split_sizes_0 = const()[name = string("op_1760_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1760_axis_0 = const()[name = string("op_1760_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1760_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1760_cast_fp16_1 = split(axis = var_1760_axis_0, split_sizes = var_1760_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1762_cast_fp16 = mul(x = var_1760_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1762_cast_fp16")];
            int32 var_1764 = const()[name = string("op_1764"), val = int32(-2)];
            bool var_1765_interleave_0 = const()[name = string("op_1765_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1765_cast_fp16 = concat(axis = var_1764, interleave = var_1765_interleave_0, values = (var_1762_cast_fp16, var_1760_cast_fp16_0))[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1766_cast_fp16 = mul(x = var_1765_cast_fp16, y = var_343_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1759_cast_fp16, y = var_1766_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1772_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_336_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [2]> var_1773_split_sizes_0 = const()[name = string("op_1773_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1773_axis_0 = const()[name = string("op_1773_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1773_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1773_cast_fp16_1 = split(axis = var_1773_axis_0, split_sizes = var_1773_split_sizes_0, x = var_1748_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1775_cast_fp16 = mul(x = var_1773_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1775_cast_fp16")];
            int32 var_1777 = const()[name = string("op_1777"), val = int32(-2)];
            bool var_1778_interleave_0 = const()[name = string("op_1778_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1778_cast_fp16 = concat(axis = var_1777, interleave = var_1778_interleave_0, values = (var_1775_cast_fp16, var_1773_cast_fp16_0))[name = string("op_1778_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1779_cast_fp16 = mul(x = var_1778_cast_fp16, y = var_343_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1779_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_6)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_8_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_8 = read_state(input = key_cache)[name = string("coreml_update_state_8")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1755_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_7)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_9_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_9 = read_state(input = value_cache)[name = string("coreml_update_state_9")];
            tensor<int32, [4]> var_1849_begin_0 = const()[name = string("op_1849_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1849_end_0 = const()[name = string("op_1849_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1849_end_mask_0 = const()[name = string("op_1849_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1849_cast_fp16 = slice_by_index(begin = var_1849_begin_0, end = var_1849_end_0, end_mask = var_1849_end_mask_0, x = coreml_update_state_8)[name = string("op_1849_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1852_axis_0 = const()[name = string("op_1852_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1852_cast_fp16_1 = split(axis = var_1852_axis_0, split_sizes = tile_8, x = var_1849_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> var_1859_begin_0 = const()[name = string("op_1859_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1859_end_0 = const()[name = string("op_1859_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1859_end_mask_0 = const()[name = string("op_1859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = coreml_update_state_9)[name = string("op_1859_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1862_axis_0 = const()[name = string("op_1862_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1862_cast_fp16_1 = split(axis = var_1862_axis_0, split_sizes = tile_9, x = var_1859_cast_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [2]> var_1865_split_sizes_0 = const()[name = string("op_1865_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1865_axis_0 = const()[name = string("op_1865_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1865_0, tensor<fp16, [1, 8, 128, 1]> var_1865_1 = split(axis = var_1865_axis_0, split_sizes = var_1865_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1865")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1852_cast_fp16_0, y = var_1865_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1868_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1872 = const()[name = string("op_1872"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1872, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1878_transpose_x_1 = const()[name = string("op_1878_transpose_x_1"), val = bool(true)];
            bool var_1878_transpose_y_1 = const()[name = string("op_1878_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1878_cast_fp16 = matmul(transpose_x = var_1878_transpose_x_1, transpose_y = var_1878_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1862_cast_fp16_0)[name = string("op_1878_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1852_cast_fp16_1, y = var_1865_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1880_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1884 = const()[name = string("op_1884"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_1884, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1862_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1892 = const()[name = string("op_1892"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_35_cast_fp16 = concat(axis = var_1892, interleave = attn_output_35_interleave_0, values = (var_1878_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1896_perm_0 = const()[name = string("op_1896_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1896_cast_fp16 = transpose(perm = var_1896_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1896_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1929_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1929_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_37_cast_fp16 = concat(axis = var_1927, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1929_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(577639808)))];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1939_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1950_split_sizes_0 = const()[name = string("op_1950_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1950_axis_0 = const()[name = string("op_1950_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1950_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1950_cast_fp16_1 = split(axis = var_1950_axis_0, split_sizes = var_1950_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1967_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<int32, [2]> var_1973_strides_0 = const()[name = string("op_1973_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1973_pad_type_0 = const()[name = string("op_1973_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1973_pad_0 = const()[name = string("op_1973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1973_dilations_0 = const()[name = string("op_1973_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1973_groups_0 = const()[name = string("op_1973_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1973_cast_fp16 = conv(dilations = var_1973_dilations_0, groups = var_1973_groups_0, pad = var_1973_pad_0, pad_type = var_1973_pad_type_0, strides = var_1973_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1950_cast_fp16_0)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_1973_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1991_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1991_cast_fp16")];
            int32 var_1989 = const()[name = string("op_1989"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_41_cast_fp16 = concat(axis = var_1989, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1991_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(577648064)))];
            fp16 var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_2001_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2012_split_sizes_0 = const()[name = string("op_2012_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2012_axis_0 = const()[name = string("op_2012_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2012_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2012_cast_fp16_1 = split(axis = var_2012_axis_0, split_sizes = var_2012_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2012_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577656320)))];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586044992)))];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = var_2012_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_cast_fp16, x = var_2012_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2069_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2076_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2080_cast_fp16 = mul(x = x_51_cast_fp16, y = var_336_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [2]> var_2081_split_sizes_0 = const()[name = string("op_2081_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2081_axis_0 = const()[name = string("op_2081_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2081_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2081_cast_fp16_1 = split(axis = var_2081_axis_0, split_sizes = var_2081_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2081_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2083_cast_fp16 = mul(x = var_2081_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2083_cast_fp16")];
            int32 var_2085 = const()[name = string("op_2085"), val = int32(-2)];
            bool var_2086_interleave_0 = const()[name = string("op_2086_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2086_cast_fp16 = concat(axis = var_2085, interleave = var_2086_interleave_0, values = (var_2083_cast_fp16, var_2081_cast_fp16_0))[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2087_cast_fp16 = mul(x = var_2086_cast_fp16, y = var_343_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2087_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2093_cast_fp16 = mul(x = var_2069_cast_fp16, y = var_336_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [2]> var_2094_split_sizes_0 = const()[name = string("op_2094_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2094_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2094_cast_fp16_1 = split(axis = var_2094_axis_0, split_sizes = var_2094_split_sizes_0, x = var_2069_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2096_cast_fp16 = mul(x = var_2094_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2096_cast_fp16")];
            int32 var_2098 = const()[name = string("op_2098"), val = int32(-2)];
            bool var_2099_interleave_0 = const()[name = string("op_2099_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2099_cast_fp16 = concat(axis = var_2098, interleave = var_2099_interleave_0, values = (var_2096_cast_fp16, var_2094_cast_fp16_0))[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2100_cast_fp16 = mul(x = var_2099_cast_fp16, y = var_343_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2093_cast_fp16, y = var_2100_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_8)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_10_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_10 = read_state(input = key_cache)[name = string("coreml_update_state_10")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2076_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_9)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_11_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_11 = read_state(input = value_cache)[name = string("coreml_update_state_11")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_10)[name = string("op_2170_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2173_axis_0 = const()[name = string("op_2173_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2173_cast_fp16_1 = split(axis = var_2173_axis_0, split_sizes = tile_10, x = var_2170_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_11)[name = string("op_2180_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2183_axis_0 = const()[name = string("op_2183_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2183_cast_fp16_1 = split(axis = var_2183_axis_0, split_sizes = tile_11, x = var_2180_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [2]> var_2186_split_sizes_0 = const()[name = string("op_2186_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2186_axis_0 = const()[name = string("op_2186_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2186_0, tensor<fp16, [1, 8, 128, 1]> var_2186_1 = split(axis = var_2186_axis_0, split_sizes = var_2186_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2186")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2173_cast_fp16_0, y = var_2186_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2189_to_fp16 = const()[name = string("op_2189_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2189_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2193 = const()[name = string("op_2193"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2193, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2199_transpose_x_1 = const()[name = string("op_2199_transpose_x_1"), val = bool(true)];
            bool var_2199_transpose_y_1 = const()[name = string("op_2199_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_1, transpose_y = var_2199_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2183_cast_fp16_0)[name = string("op_2199_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2173_cast_fp16_1, y = var_2186_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2205 = const()[name = string("op_2205"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2205, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2183_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_43_cast_fp16 = concat(axis = var_2213, interleave = attn_output_43_interleave_0, values = (var_2199_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2217_perm_0 = const()[name = string("op_2217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2217_cast_fp16 = transpose(perm = var_2217_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2217_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2250_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_45_cast_fp16 = concat(axis = var_2248, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2250_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(587093632)))];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2260_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2271_split_sizes_0 = const()[name = string("op_2271_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2271_axis_0 = const()[name = string("op_2271_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2271_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2271_cast_fp16_1 = split(axis = var_2271_axis_0, split_sizes = var_2271_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587101888)))];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = var_2271_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2288_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2288_cast_fp16")];
            tensor<int32, [2]> var_2294_strides_0 = const()[name = string("op_2294_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2294_pad_type_0 = const()[name = string("op_2294_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2294_pad_0 = const()[name = string("op_2294_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2294_dilations_0 = const()[name = string("op_2294_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2294_groups_0 = const()[name = string("op_2294_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2294_cast_fp16 = conv(dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2271_cast_fp16_0)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_2294_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2312_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_49_cast_fp16 = concat(axis = var_2310, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2312_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(612267776)))];
            fp16 var_2322_to_fp16 = const()[name = string("op_2322_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2322_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = string("op_2333_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2333_axis_0 = const()[name = string("op_2333_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2333_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612276032)))];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620664704)))];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = var_2333_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_cast_fp16, x = var_2333_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2390_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2397_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2397_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2401_cast_fp16 = mul(x = x_61_cast_fp16, y = var_336_cast_fp16)[name = string("op_2401_cast_fp16")];
            tensor<int32, [2]> var_2402_split_sizes_0 = const()[name = string("op_2402_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2402_axis_0 = const()[name = string("op_2402_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2402_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2402_cast_fp16_1 = split(axis = var_2402_axis_0, split_sizes = var_2402_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2402_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2404_cast_fp16 = mul(x = var_2402_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2404_cast_fp16")];
            int32 var_2406 = const()[name = string("op_2406"), val = int32(-2)];
            bool var_2407_interleave_0 = const()[name = string("op_2407_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2407_cast_fp16 = concat(axis = var_2406, interleave = var_2407_interleave_0, values = (var_2404_cast_fp16, var_2402_cast_fp16_0))[name = string("op_2407_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2408_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_343_cast_fp16)[name = string("op_2408_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2408_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2414_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_336_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [2]> var_2415_split_sizes_0 = const()[name = string("op_2415_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2415_axis_0 = const()[name = string("op_2415_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2415_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2415_cast_fp16_1 = split(axis = var_2415_axis_0, split_sizes = var_2415_split_sizes_0, x = var_2390_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2417_cast_fp16 = mul(x = var_2415_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2417_cast_fp16")];
            int32 var_2419 = const()[name = string("op_2419"), val = int32(-2)];
            bool var_2420_interleave_0 = const()[name = string("op_2420_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2420_cast_fp16 = concat(axis = var_2419, interleave = var_2420_interleave_0, values = (var_2417_cast_fp16, var_2415_cast_fp16_0))[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2421_cast_fp16 = mul(x = var_2420_cast_fp16, y = var_343_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2421_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_10)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_12_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_12 = read_state(input = key_cache)[name = string("coreml_update_state_12")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2397_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_11)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_13_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_13 = read_state(input = value_cache)[name = string("coreml_update_state_13")];
            tensor<int32, [4]> var_2491_begin_0 = const()[name = string("op_2491_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2491_end_0 = const()[name = string("op_2491_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2491_end_mask_0 = const()[name = string("op_2491_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2491_cast_fp16 = slice_by_index(begin = var_2491_begin_0, end = var_2491_end_0, end_mask = var_2491_end_mask_0, x = coreml_update_state_12)[name = string("op_2491_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2494_axis_0 = const()[name = string("op_2494_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2494_cast_fp16_1 = split(axis = var_2494_axis_0, split_sizes = tile_12, x = var_2491_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> var_2501_begin_0 = const()[name = string("op_2501_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2501_end_0 = const()[name = string("op_2501_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2501_end_mask_0 = const()[name = string("op_2501_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2501_cast_fp16 = slice_by_index(begin = var_2501_begin_0, end = var_2501_end_0, end_mask = var_2501_end_mask_0, x = coreml_update_state_13)[name = string("op_2501_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2504_axis_0 = const()[name = string("op_2504_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2504_cast_fp16_1 = split(axis = var_2504_axis_0, split_sizes = tile_13, x = var_2501_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2507_0, tensor<fp16, [1, 8, 128, 1]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2507")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2494_cast_fp16_0, y = var_2507_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2510_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2514 = const()[name = string("op_2514"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2514, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2520_transpose_x_1 = const()[name = string("op_2520_transpose_x_1"), val = bool(true)];
            bool var_2520_transpose_y_1 = const()[name = string("op_2520_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_1, transpose_y = var_2520_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2504_cast_fp16_0)[name = string("op_2520_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2494_cast_fp16_1, y = var_2507_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2522_to_fp16 = const()[name = string("op_2522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2522_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2526 = const()[name = string("op_2526"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2526, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2504_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2534 = const()[name = string("op_2534"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_51_cast_fp16 = concat(axis = var_2534, interleave = attn_output_51_interleave_0, values = (var_2520_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2538_perm_0 = const()[name = string("op_2538_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2538_cast_fp16 = transpose(perm = var_2538_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2538_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2571_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2571_cast_fp16")];
            int32 var_2569 = const()[name = string("op_2569"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_53_cast_fp16 = concat(axis = var_2569, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2571_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(621713344)))];
            fp16 var_2581_to_fp16 = const()[name = string("op_2581_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2581_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2592_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2609_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [2]> var_2615_strides_0 = const()[name = string("op_2615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2615_pad_type_0 = const()[name = string("op_2615_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2615_pad_0 = const()[name = string("op_2615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2615_dilations_0 = const()[name = string("op_2615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2615_groups_0 = const()[name = string("op_2615_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2615_cast_fp16 = conv(dilations = var_2615_dilations_0, groups = var_2615_groups_0, pad = var_2615_pad_0, pad_type = var_2615_pad_type_0, strides = var_2615_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2592_cast_fp16_0)[name = string("op_2615_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_2615_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_cast_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2633_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2633_cast_fp16")];
            int32 var_2631 = const()[name = string("op_2631"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_57_cast_fp16 = concat(axis = var_2631, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2633_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(621721600)))];
            fp16 var_2643_to_fp16 = const()[name = string("op_2643_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2643_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2654_split_sizes_0 = const()[name = string("op_2654_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2654_axis_0 = const()[name = string("op_2654_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2654_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2654_cast_fp16_1 = split(axis = var_2654_axis_0, split_sizes = var_2654_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621729856)))];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630118528)))];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = var_2654_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_cast_fp16, x = var_2654_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2711_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2718_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_336_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<int32, [2]> var_2723_split_sizes_0 = const()[name = string("op_2723_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2723_axis_0 = const()[name = string("op_2723_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2723_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2723_cast_fp16_1 = split(axis = var_2723_axis_0, split_sizes = var_2723_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2725_cast_fp16 = mul(x = var_2723_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2725_cast_fp16")];
            int32 var_2727 = const()[name = string("op_2727"), val = int32(-2)];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2728_cast_fp16 = concat(axis = var_2727, interleave = var_2728_interleave_0, values = (var_2725_cast_fp16, var_2723_cast_fp16_0))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2729_cast_fp16 = mul(x = var_2728_cast_fp16, y = var_343_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2722_cast_fp16, y = var_2729_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2735_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_336_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [2]> var_2736_split_sizes_0 = const()[name = string("op_2736_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2736_axis_0 = const()[name = string("op_2736_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2736_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2736_cast_fp16_1 = split(axis = var_2736_axis_0, split_sizes = var_2736_split_sizes_0, x = var_2711_cast_fp16)[name = string("op_2736_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2738_cast_fp16 = mul(x = var_2736_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-2)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2738_cast_fp16, var_2736_cast_fp16_0))[name = string("op_2741_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2742_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_343_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2742_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_12)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_14_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_14 = read_state(input = key_cache)[name = string("coreml_update_state_14")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2718_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_13)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_15_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_15 = read_state(input = value_cache)[name = string("coreml_update_state_15")];
            tensor<int32, [4]> var_2812_begin_0 = const()[name = string("op_2812_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2812_end_0 = const()[name = string("op_2812_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = string("op_2812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = coreml_update_state_14)[name = string("op_2812_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2815_axis_0 = const()[name = string("op_2815_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2815_cast_fp16_1 = split(axis = var_2815_axis_0, split_sizes = tile_14, x = var_2812_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<int32, [4]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = coreml_update_state_15)[name = string("op_2822_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2825_axis_0 = const()[name = string("op_2825_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2825_cast_fp16_1 = split(axis = var_2825_axis_0, split_sizes = tile_15, x = var_2822_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<int32, [2]> var_2828_split_sizes_0 = const()[name = string("op_2828_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2828_axis_0 = const()[name = string("op_2828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2828_0, tensor<fp16, [1, 8, 128, 1]> var_2828_1 = split(axis = var_2828_axis_0, split_sizes = var_2828_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2828")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2815_cast_fp16_0, y = var_2828_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2831_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2835 = const()[name = string("op_2835"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2835, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2841_transpose_x_1 = const()[name = string("op_2841_transpose_x_1"), val = bool(true)];
            bool var_2841_transpose_y_1 = const()[name = string("op_2841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2841_cast_fp16 = matmul(transpose_x = var_2841_transpose_x_1, transpose_y = var_2841_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2825_cast_fp16_0)[name = string("op_2841_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2815_cast_fp16_1, y = var_2828_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2843_to_fp16 = const()[name = string("op_2843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2843_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2847 = const()[name = string("op_2847"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2847, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2825_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2855 = const()[name = string("op_2855"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_59_cast_fp16 = concat(axis = var_2855, interleave = attn_output_59_interleave_0, values = (var_2841_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2859_perm_0 = const()[name = string("op_2859_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2859_cast_fp16 = transpose(perm = var_2859_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2859_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2892_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
            int32 var_2890 = const()[name = string("op_2890"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_61_cast_fp16 = concat(axis = var_2890, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2892_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(631167168)))];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2902_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2913_split_sizes_0 = const()[name = string("op_2913_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2913_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2913_cast_fp16_1 = split(axis = var_2913_axis_0, split_sizes = var_2913_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2913_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2930_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2930_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631175424)))];
            tensor<int32, [2]> var_2936_strides_0 = const()[name = string("op_2936_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2936_pad_type_0 = const()[name = string("op_2936_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2936_pad_0 = const()[name = string("op_2936_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2936_dilations_0 = const()[name = string("op_2936_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2936_groups_0 = const()[name = string("op_2936_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2936_cast_fp16 = conv(dilations = var_2936_dilations_0, groups = var_2936_groups_0, pad = var_2936_pad_0, pad_type = var_2936_pad_type_0, strides = var_2936_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2913_cast_fp16_0)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2936_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656341312)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2954_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2954_cast_fp16")];
            int32 var_2952 = const()[name = string("op_2952"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_65_cast_fp16 = concat(axis = var_2952, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2954_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(681507200)))];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2964_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2975_split_sizes_0 = const()[name = string("op_2975_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2975_axis_0 = const()[name = string("op_2975_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2975_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2975_cast_fp16_1 = split(axis = var_2975_axis_0, split_sizes = var_2975_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681515456)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689904128)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2975_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_cast_fp16, x = var_2975_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3032_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3039_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3043_cast_fp16 = mul(x = x_81_cast_fp16, y = var_336_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<int32, [2]> var_3044_split_sizes_0 = const()[name = string("op_3044_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3044_axis_0 = const()[name = string("op_3044_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3044_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3044_cast_fp16_1 = split(axis = var_3044_axis_0, split_sizes = var_3044_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3046_cast_fp16 = mul(x = var_3044_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            int32 var_3048 = const()[name = string("op_3048"), val = int32(-2)];
            bool var_3049_interleave_0 = const()[name = string("op_3049_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3049_cast_fp16 = concat(axis = var_3048, interleave = var_3049_interleave_0, values = (var_3046_cast_fp16, var_3044_cast_fp16_0))[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3050_cast_fp16 = mul(x = var_3049_cast_fp16, y = var_343_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3050_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3056_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_336_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<int32, [2]> var_3057_split_sizes_0 = const()[name = string("op_3057_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3057_axis_0 = const()[name = string("op_3057_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3057_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3057_cast_fp16_1 = split(axis = var_3057_axis_0, split_sizes = var_3057_split_sizes_0, x = var_3032_cast_fp16)[name = string("op_3057_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3059_cast_fp16 = mul(x = var_3057_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            int32 var_3061 = const()[name = string("op_3061"), val = int32(-2)];
            bool var_3062_interleave_0 = const()[name = string("op_3062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3062_cast_fp16 = concat(axis = var_3061, interleave = var_3062_interleave_0, values = (var_3059_cast_fp16, var_3057_cast_fp16_0))[name = string("op_3062_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3063_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_343_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3063_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_14)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_16_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_16 = read_state(input = key_cache)[name = string("coreml_update_state_16")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3039_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_15)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_17_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_17 = read_state(input = value_cache)[name = string("coreml_update_state_17")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, x = coreml_update_state_16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3136_axis_0 = const()[name = string("op_3136_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3136_cast_fp16_1 = split(axis = var_3136_axis_0, split_sizes = tile_16, x = var_3133_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = coreml_update_state_17)[name = string("op_3143_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = tile_17, x = var_3143_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<int32, [2]> var_3149_split_sizes_0 = const()[name = string("op_3149_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3149_axis_0 = const()[name = string("op_3149_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3149_0, tensor<fp16, [1, 8, 128, 1]> var_3149_1 = split(axis = var_3149_axis_0, split_sizes = var_3149_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3149")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3136_cast_fp16_0, y = var_3149_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3152_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3156 = const()[name = string("op_3156"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3156, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3162_transpose_x_1 = const()[name = string("op_3162_transpose_x_1"), val = bool(true)];
            bool var_3162_transpose_y_1 = const()[name = string("op_3162_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3162_cast_fp16 = matmul(transpose_x = var_3162_transpose_x_1, transpose_y = var_3162_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3146_cast_fp16_0)[name = string("op_3162_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3136_cast_fp16_1, y = var_3149_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3164_to_fp16 = const()[name = string("op_3164_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3164_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3168 = const()[name = string("op_3168"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_143_cast_fp16 = softmax(axis = var_3168, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_143_cast_fp16, y = var_3146_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3176 = const()[name = string("op_3176"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_67_cast_fp16 = concat(axis = var_3176, interleave = attn_output_67_interleave_0, values = (var_3162_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3180_perm_0 = const()[name = string("op_3180_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3180_cast_fp16 = transpose(perm = var_3180_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3180_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_cast_fp16, x = attn_output_71_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3213_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3213_cast_fp16")];
            int32 var_3211 = const()[name = string("op_3211"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_69_cast_fp16 = concat(axis = var_3211, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3213_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(690952768)))];
            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3223_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3234_split_sizes_0 = const()[name = string("op_3234_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3234_axis_0 = const()[name = string("op_3234_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3234_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3234_cast_fp16_1 = split(axis = var_3234_axis_0, split_sizes = var_3234_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3251_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [2]> var_3257_strides_0 = const()[name = string("op_3257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3257_pad_type_0 = const()[name = string("op_3257_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3257_pad_0 = const()[name = string("op_3257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3257_dilations_0 = const()[name = string("op_3257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3257_groups_0 = const()[name = string("op_3257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3257_cast_fp16 = conv(dilations = var_3257_dilations_0, groups = var_3257_groups_0, pad = var_3257_pad_0, pad_type = var_3257_pad_type_0, strides = var_3257_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3234_cast_fp16_0)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_89_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_3257_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3275_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3275_cast_fp16")];
            int32 var_3273 = const()[name = string("op_3273"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_73_cast_fp16 = concat(axis = var_3273, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3275_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(690961024)))];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3285_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3296_split_sizes_0 = const()[name = string("op_3296_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3296_axis_0 = const()[name = string("op_3296_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3296_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3296_cast_fp16_1 = split(axis = var_3296_axis_0, split_sizes = var_3296_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690969280)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699357952)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3296_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3296_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3353_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3360_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3364_cast_fp16 = mul(x = x_91_cast_fp16, y = var_336_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [2]> var_3365_split_sizes_0 = const()[name = string("op_3365_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3365_axis_0 = const()[name = string("op_3365_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3365_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3365_cast_fp16_1 = split(axis = var_3365_axis_0, split_sizes = var_3365_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3367_cast_fp16 = mul(x = var_3365_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
            int32 var_3369 = const()[name = string("op_3369"), val = int32(-2)];
            bool var_3370_interleave_0 = const()[name = string("op_3370_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3370_cast_fp16 = concat(axis = var_3369, interleave = var_3370_interleave_0, values = (var_3367_cast_fp16, var_3365_cast_fp16_0))[name = string("op_3370_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3371_cast_fp16 = mul(x = var_3370_cast_fp16, y = var_343_cast_fp16)[name = string("op_3371_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_57_cast_fp16 = add(x = var_3364_cast_fp16, y = var_3371_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3377_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_336_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [2]> var_3378_split_sizes_0 = const()[name = string("op_3378_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3378_axis_0 = const()[name = string("op_3378_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3378_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3378_cast_fp16_1 = split(axis = var_3378_axis_0, split_sizes = var_3378_split_sizes_0, x = var_3353_cast_fp16)[name = string("op_3378_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3380_cast_fp16 = mul(x = var_3378_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3380_cast_fp16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(-2)];
            bool var_3383_interleave_0 = const()[name = string("op_3383_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3383_cast_fp16 = concat(axis = var_3382, interleave = var_3383_interleave_0, values = (var_3380_cast_fp16, var_3378_cast_fp16_0))[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3384_cast_fp16 = mul(x = var_3383_cast_fp16, y = var_343_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_95_cast_fp16 = add(x = var_3377_cast_fp16, y = var_3384_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [10, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_16)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_18_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_18 = read_state(input = key_cache)[name = string("coreml_update_state_18")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3360_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [10, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_17)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_19_write_state")];
            tensor<fp16, [10, 2, 2048, 128]> coreml_update_state_19 = read_state(input = value_cache)[name = string("coreml_update_state_19")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = string("op_3454_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = string("op_3454_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = string("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = coreml_update_state_18)[name = string("op_3454_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3457_axis_0 = const()[name = string("op_3457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3457_cast_fp16_1 = split(axis = var_3457_axis_0, split_sizes = tile_18, x = var_3454_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [4]> var_3464_begin_0 = const()[name = string("op_3464_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3464_end_0 = const()[name = string("op_3464_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3464_end_mask_0 = const()[name = string("op_3464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3464_cast_fp16 = slice_by_index(begin = var_3464_begin_0, end = var_3464_end_0, end_mask = var_3464_end_mask_0, x = coreml_update_state_19)[name = string("op_3464_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3467_axis_0 = const()[name = string("op_3467_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3467_cast_fp16_1 = split(axis = var_3467_axis_0, split_sizes = tile_19, x = var_3464_cast_fp16)[name = string("op_3467_cast_fp16")];
            tensor<int32, [2]> var_3470_split_sizes_0 = const()[name = string("op_3470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3470_axis_0 = const()[name = string("op_3470_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3470_0, tensor<fp16, [1, 8, 128, 1]> var_3470_1 = split(axis = var_3470_axis_0, split_sizes = var_3470_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3470")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3457_cast_fp16_0, y = var_3470_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3473_to_fp16 = const()[name = string("op_3473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3473_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3477 = const()[name = string("op_3477"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_151_cast_fp16 = softmax(axis = var_3477, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3483_transpose_x_1 = const()[name = string("op_3483_transpose_x_1"), val = bool(true)];
            bool var_3483_transpose_y_1 = const()[name = string("op_3483_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3467_cast_fp16_0)[name = string("op_3483_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3457_cast_fp16_1, y = var_3470_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3485_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3489 = const()[name = string("op_3489"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_3489, x = attn_weights_157_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_cast_fp16, y = var_3467_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3497 = const()[name = string("op_3497"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_75_cast_fp16 = concat(axis = var_3497, interleave = attn_output_75_interleave_0, values = (var_3483_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3501_perm_0 = const()[name = string("op_3501_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3501_cast_fp16 = transpose(perm = var_3501_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_119x, x = var_3501_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3534_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3534_cast_fp16")];
            int32 var_3532 = const()[name = string("op_3532"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_77_cast_fp16 = concat(axis = var_3532, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3534_cast_fp16))[name = string("doubled_77_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(700406592)))];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3544_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3555_split_sizes_0 = const()[name = string("op_3555_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3555_axis_0 = const()[name = string("op_3555_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3555_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3555_cast_fp16_1 = split(axis = var_3555_axis_0, split_sizes = var_3555_split_sizes_0, x = out_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3572_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3572_cast_fp16")];
            tensor<int32, [2]> var_3578_strides_0 = const()[name = string("op_3578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3578_pad_type_0 = const()[name = string("op_3578_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3578_pad_0 = const()[name = string("op_3578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3578_dilations_0 = const()[name = string("op_3578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3578_groups_0 = const()[name = string("op_3578_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3578_cast_fp16 = conv(dilations = var_3578_dilations_0, groups = var_3578_groups_0, pad = var_3578_pad_0, pad_type = var_3578_pad_type_0, strides = var_3578_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3555_cast_fp16_0)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_3578_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states = add(x = hidden_states_95_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_3587_cast_fp16")];
        } -> (hidden_states);
}