program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-milinternal", ""}, {"coremltools-version", "9.0"}})]
{
    func length_1<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(1)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 1, 128, 1]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 1, 128, 1]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 1, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_5")];
            tensor<fp16, [1, 1, 1, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_87")];
            tensor<bool, [1, 1, 2048, 1]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_28_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_28 = read_state(input = key_cache)[name = string("coreml_update_state_28")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_29_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_29 = read_state(input = value_cache)[name = string("coreml_update_state_29")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_28)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_29)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_698_0, tensor<fp16, [1, 8, 128, 1]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_28)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_30_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_30 = read_state(input = key_cache)[name = string("coreml_update_state_30")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_29)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_31_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_31 = read_state(input = value_cache)[name = string("coreml_update_state_31")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_30)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_31)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1019_0, tensor<fp16, [1, 8, 128, 1]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_30)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_32_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_32 = read_state(input = key_cache)[name = string("coreml_update_state_32")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_31)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_33_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_33 = read_state(input = value_cache)[name = string("coreml_update_state_33")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_32)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_33)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1340_0, tensor<fp16, [1, 8, 128, 1]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_32)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_34_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_34 = read_state(input = key_cache)[name = string("coreml_update_state_34")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_33)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_35_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_35 = read_state(input = value_cache)[name = string("coreml_update_state_35")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_34)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_35)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1661_0, tensor<fp16, [1, 8, 128, 1]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_34)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_36_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_36 = read_state(input = key_cache)[name = string("coreml_update_state_36")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_35)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_37_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_37 = read_state(input = value_cache)[name = string("coreml_update_state_37")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_36)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_37)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1982_0, tensor<fp16, [1, 8, 128, 1]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_36)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_38_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_38 = read_state(input = key_cache)[name = string("coreml_update_state_38")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_37)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_39_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_39 = read_state(input = value_cache)[name = string("coreml_update_state_39")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_38)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_39)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2303_0, tensor<fp16, [1, 8, 128, 1]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_38)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_40_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_40 = read_state(input = key_cache)[name = string("coreml_update_state_40")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_39)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_41_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_41 = read_state(input = value_cache)[name = string("coreml_update_state_41")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_40)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_41)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2624_0, tensor<fp16, [1, 8, 128, 1]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_40)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_42_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_42 = read_state(input = key_cache)[name = string("coreml_update_state_42")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_41)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_43_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_43 = read_state(input = value_cache)[name = string("coreml_update_state_43")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_42)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_43)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2945_0, tensor<fp16, [1, 8, 128, 1]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_42)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_44_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_44 = read_state(input = key_cache)[name = string("coreml_update_state_44")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_43)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_45_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_45 = read_state(input = value_cache)[name = string("coreml_update_state_45")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_44)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_45)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3266_0, tensor<fp16, [1, 8, 128, 1]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_44)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_46_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_46 = read_state(input = key_cache)[name = string("coreml_update_state_46")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_45)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_47_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_47 = read_state(input = value_cache)[name = string("coreml_update_state_47")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_46)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_47)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3587_0, tensor<fp16, [1, 8, 128, 1]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_46)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_48_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_48 = read_state(input = key_cache)[name = string("coreml_update_state_48")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_47)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_49_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_49 = read_state(input = value_cache)[name = string("coreml_update_state_49")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_48)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_49)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3908_0, tensor<fp16, [1, 8, 128, 1]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_48)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_50_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_50 = read_state(input = key_cache)[name = string("coreml_update_state_50")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_49)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_51_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_51 = read_state(input = value_cache)[name = string("coreml_update_state_51")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_50)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_51)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4229_0, tensor<fp16, [1, 8, 128, 1]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_50)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_52_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_52 = read_state(input = key_cache)[name = string("coreml_update_state_52")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_51)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_53_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_53 = read_state(input = value_cache)[name = string("coreml_update_state_53")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_52)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_53)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4550_0, tensor<fp16, [1, 8, 128, 1]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_52)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_54_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_54 = read_state(input = key_cache)[name = string("coreml_update_state_54")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_53)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_55_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_55 = read_state(input = value_cache)[name = string("coreml_update_state_55")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_54)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_55)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4871_0, tensor<fp16, [1, 8, 128, 1]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states, tensor<fp16, [1, 2048, 1, 1]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
    func length_128<ios18>(tensor<fp32, [1, 2048, 1, 128]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [128]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(128)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [128]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 128]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_314")];
            tensor<fp16, [1, 1, 128, 128]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_313")];
            tensor<fp16, [1, 1, 128, 128]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 128, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 128, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_25")];
            tensor<fp16, [1, 1, 128, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 128]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_312")];
            tensor<bool, [1, 1, 2048, 128]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 128]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_311")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_168_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_168 = read_state(input = key_cache)[name = string("coreml_update_state_168")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_310")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_169_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_169 = read_state(input = value_cache)[name = string("coreml_update_state_169")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_168)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_169)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_698_0, tensor<fp16, [1, 8, 128, 128]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_309")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_308")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_168)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_170_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_170 = read_state(input = key_cache)[name = string("coreml_update_state_170")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_307")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_169)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_171_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_171 = read_state(input = value_cache)[name = string("coreml_update_state_171")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_170)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_171)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1019_0, tensor<fp16, [1, 8, 128, 128]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_306")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_305")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_170)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_172_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_172 = read_state(input = key_cache)[name = string("coreml_update_state_172")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_304")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_171)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_173_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_173 = read_state(input = value_cache)[name = string("coreml_update_state_173")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_172)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_173)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1340_0, tensor<fp16, [1, 8, 128, 128]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_303")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_302")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_172)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_174_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_174 = read_state(input = key_cache)[name = string("coreml_update_state_174")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_301")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_173)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_175_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_175 = read_state(input = value_cache)[name = string("coreml_update_state_175")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_174)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_175)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1661_0, tensor<fp16, [1, 8, 128, 128]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_300")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_299")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_174)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_176_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_176 = read_state(input = key_cache)[name = string("coreml_update_state_176")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_298")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_175)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_177_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_177 = read_state(input = value_cache)[name = string("coreml_update_state_177")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_176)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_177)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1982_0, tensor<fp16, [1, 8, 128, 128]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_297")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_296")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_176)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_178_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_178 = read_state(input = key_cache)[name = string("coreml_update_state_178")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_295")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_177)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_179_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_179 = read_state(input = value_cache)[name = string("coreml_update_state_179")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_178)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_179)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2303_0, tensor<fp16, [1, 8, 128, 128]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_294")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_293")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_178)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_180_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_180 = read_state(input = key_cache)[name = string("coreml_update_state_180")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_292")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_179)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_181_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_181 = read_state(input = value_cache)[name = string("coreml_update_state_181")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_180)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_181)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2624_0, tensor<fp16, [1, 8, 128, 128]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_291")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_290")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_180)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_182_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_182 = read_state(input = key_cache)[name = string("coreml_update_state_182")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_289")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_181)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_183_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_183 = read_state(input = value_cache)[name = string("coreml_update_state_183")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_182)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_183)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2945_0, tensor<fp16, [1, 8, 128, 128]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_288")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_287")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_182)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_184_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_184 = read_state(input = key_cache)[name = string("coreml_update_state_184")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_286")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_183)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_185_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_185 = read_state(input = value_cache)[name = string("coreml_update_state_185")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_184)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_185)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3266_0, tensor<fp16, [1, 8, 128, 128]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_285")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_284")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_184)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_186_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_186 = read_state(input = key_cache)[name = string("coreml_update_state_186")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_283")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_185)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_187_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_187 = read_state(input = value_cache)[name = string("coreml_update_state_187")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_186)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_187)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3587_0, tensor<fp16, [1, 8, 128, 128]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_282")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_281")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_186)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_188_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_188 = read_state(input = key_cache)[name = string("coreml_update_state_188")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_280")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_187)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_189_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_189 = read_state(input = value_cache)[name = string("coreml_update_state_189")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_188)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_189)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3908_0, tensor<fp16, [1, 8, 128, 128]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_279")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_278")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_188)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_190_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_190 = read_state(input = key_cache)[name = string("coreml_update_state_190")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_277")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_189)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_191_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_191 = read_state(input = value_cache)[name = string("coreml_update_state_191")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_190)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_191)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_4229_0, tensor<fp16, [1, 8, 128, 128]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_276")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_275")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_190)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_192_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_192 = read_state(input = key_cache)[name = string("coreml_update_state_192")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_274")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_191)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_193_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_193 = read_state(input = value_cache)[name = string("coreml_update_state_193")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_192)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_193)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_4550_0, tensor<fp16, [1, 8, 128, 128]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_273")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_272")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_192)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_194_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_194 = read_state(input = key_cache)[name = string("coreml_update_state_194")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_271")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_193)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_195_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_195 = read_state(input = value_cache)[name = string("coreml_update_state_195")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_194)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_195)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_4871_0, tensor<fp16, [1, 8, 128, 128]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_270")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states, tensor<fp16, [1, 2048, 1, 128]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
    func length_16<ios18>(tensor<fp32, [1, 2048, 1, 16]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [16]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(16)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [16]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 16]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 1, 128, 16]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 1, 128, 16]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 16, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 16, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 16, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_13")];
            tensor<fp16, [1, 1, 16, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 16]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_177")];
            tensor<bool, [1, 1, 2048, 16]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 16]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_84 = read_state(input = key_cache)[name = string("coreml_update_state_84")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_85 = read_state(input = value_cache)[name = string("coreml_update_state_85")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_84)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_85)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_698_0, tensor<fp16, [1, 8, 128, 16]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_84)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_86 = read_state(input = key_cache)[name = string("coreml_update_state_86")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_85)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_87 = read_state(input = value_cache)[name = string("coreml_update_state_87")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_86)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_87)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1019_0, tensor<fp16, [1, 8, 128, 16]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_86)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_88 = read_state(input = key_cache)[name = string("coreml_update_state_88")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_87)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_89 = read_state(input = value_cache)[name = string("coreml_update_state_89")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_88)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_89)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1340_0, tensor<fp16, [1, 8, 128, 16]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_88)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_90 = read_state(input = key_cache)[name = string("coreml_update_state_90")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_89)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_91 = read_state(input = value_cache)[name = string("coreml_update_state_91")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_90)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_91)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1661_0, tensor<fp16, [1, 8, 128, 16]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_90)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_92 = read_state(input = key_cache)[name = string("coreml_update_state_92")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_91)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_93 = read_state(input = value_cache)[name = string("coreml_update_state_93")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_92)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_93)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1982_0, tensor<fp16, [1, 8, 128, 16]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_92)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_94_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_94 = read_state(input = key_cache)[name = string("coreml_update_state_94")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_93)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_95_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_95 = read_state(input = value_cache)[name = string("coreml_update_state_95")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_94)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_95)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2303_0, tensor<fp16, [1, 8, 128, 16]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_94)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_96_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_96 = read_state(input = key_cache)[name = string("coreml_update_state_96")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_95)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_97_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_97 = read_state(input = value_cache)[name = string("coreml_update_state_97")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_96)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_97)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2624_0, tensor<fp16, [1, 8, 128, 16]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_96)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_98_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_98 = read_state(input = key_cache)[name = string("coreml_update_state_98")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_97)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_99_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_99 = read_state(input = value_cache)[name = string("coreml_update_state_99")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_98)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_99)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2945_0, tensor<fp16, [1, 8, 128, 16]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_98)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_100 = read_state(input = key_cache)[name = string("coreml_update_state_100")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_99)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_101 = read_state(input = value_cache)[name = string("coreml_update_state_101")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_100)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_101)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3266_0, tensor<fp16, [1, 8, 128, 16]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_100)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_102 = read_state(input = key_cache)[name = string("coreml_update_state_102")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_101)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_103 = read_state(input = value_cache)[name = string("coreml_update_state_103")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_102)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_103)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3587_0, tensor<fp16, [1, 8, 128, 16]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_102)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_104 = read_state(input = key_cache)[name = string("coreml_update_state_104")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_103)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_105 = read_state(input = value_cache)[name = string("coreml_update_state_105")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_104)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_105)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3908_0, tensor<fp16, [1, 8, 128, 16]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_104)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_106 = read_state(input = key_cache)[name = string("coreml_update_state_106")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_105)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_107 = read_state(input = value_cache)[name = string("coreml_update_state_107")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_106)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_107)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_4229_0, tensor<fp16, [1, 8, 128, 16]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_106)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_108 = read_state(input = key_cache)[name = string("coreml_update_state_108")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_107)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_109 = read_state(input = value_cache)[name = string("coreml_update_state_109")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_108)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_109)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_4550_0, tensor<fp16, [1, 8, 128, 16]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_108)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_110 = read_state(input = key_cache)[name = string("coreml_update_state_110")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_109)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_111 = read_state(input = value_cache)[name = string("coreml_update_state_111")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_110)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_111)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_4871_0, tensor<fp16, [1, 8, 128, 16]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states, tensor<fp16, [1, 2048, 1, 16]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
    func length_32<ios18>(tensor<fp32, [1, 2048, 1, 32]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [32]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(32)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [32]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 32]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [1, 1, 128, 32]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [1, 1, 128, 32]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 32, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 32, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 32, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_17")];
            tensor<fp16, [1, 1, 32, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 32]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_222")];
            tensor<bool, [1, 1, 2048, 32]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 32]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_112_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_112 = read_state(input = key_cache)[name = string("coreml_update_state_112")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_113_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_113 = read_state(input = value_cache)[name = string("coreml_update_state_113")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_112)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_113)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_698_0, tensor<fp16, [1, 8, 128, 32]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_112)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_114_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_114 = read_state(input = key_cache)[name = string("coreml_update_state_114")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_113)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_115_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_115 = read_state(input = value_cache)[name = string("coreml_update_state_115")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_114)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_115)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1019_0, tensor<fp16, [1, 8, 128, 32]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_114)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_116_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_116 = read_state(input = key_cache)[name = string("coreml_update_state_116")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_115)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_117_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_117 = read_state(input = value_cache)[name = string("coreml_update_state_117")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_116)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_117)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1340_0, tensor<fp16, [1, 8, 128, 32]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_116)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_118_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_118 = read_state(input = key_cache)[name = string("coreml_update_state_118")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_117)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_119_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_119 = read_state(input = value_cache)[name = string("coreml_update_state_119")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_118)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_119)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1661_0, tensor<fp16, [1, 8, 128, 32]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_118)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_120_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_120 = read_state(input = key_cache)[name = string("coreml_update_state_120")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_119)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_121_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_121 = read_state(input = value_cache)[name = string("coreml_update_state_121")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_120)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_121)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1982_0, tensor<fp16, [1, 8, 128, 32]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_120)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_122_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_122 = read_state(input = key_cache)[name = string("coreml_update_state_122")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_121)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_123_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_123 = read_state(input = value_cache)[name = string("coreml_update_state_123")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_122)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_123)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2303_0, tensor<fp16, [1, 8, 128, 32]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_122)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_124_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_124 = read_state(input = key_cache)[name = string("coreml_update_state_124")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_123)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_125_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_125 = read_state(input = value_cache)[name = string("coreml_update_state_125")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_124)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_125)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2624_0, tensor<fp16, [1, 8, 128, 32]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_124)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_126_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_126 = read_state(input = key_cache)[name = string("coreml_update_state_126")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_125)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_127_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_127 = read_state(input = value_cache)[name = string("coreml_update_state_127")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_126)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_127)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2945_0, tensor<fp16, [1, 8, 128, 32]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_126)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_128_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_128 = read_state(input = key_cache)[name = string("coreml_update_state_128")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_127)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_129_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_129 = read_state(input = value_cache)[name = string("coreml_update_state_129")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_128)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_129)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3266_0, tensor<fp16, [1, 8, 128, 32]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_128)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_130_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_130 = read_state(input = key_cache)[name = string("coreml_update_state_130")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_129)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_131_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_131 = read_state(input = value_cache)[name = string("coreml_update_state_131")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_130)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_131)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3587_0, tensor<fp16, [1, 8, 128, 32]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_130)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_132_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_132 = read_state(input = key_cache)[name = string("coreml_update_state_132")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_131)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_133_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_133 = read_state(input = value_cache)[name = string("coreml_update_state_133")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_132)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_133)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3908_0, tensor<fp16, [1, 8, 128, 32]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_132)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_134_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_134 = read_state(input = key_cache)[name = string("coreml_update_state_134")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_133)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_135_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_135 = read_state(input = value_cache)[name = string("coreml_update_state_135")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_134)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_135)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_4229_0, tensor<fp16, [1, 8, 128, 32]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_134)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_136_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_136 = read_state(input = key_cache)[name = string("coreml_update_state_136")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_135)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_137_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_137 = read_state(input = value_cache)[name = string("coreml_update_state_137")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_136)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_137)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_4550_0, tensor<fp16, [1, 8, 128, 32]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_136)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_138_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_138 = read_state(input = key_cache)[name = string("coreml_update_state_138")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_137)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_139_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_139 = read_state(input = value_cache)[name = string("coreml_update_state_139")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_138)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_139)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_4871_0, tensor<fp16, [1, 8, 128, 32]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states, tensor<fp16, [1, 2048, 1, 32]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
    func length_64<ios18>(tensor<fp32, [1, 2048, 1, 64]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [64]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(64)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [64]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 64]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_269")];
            tensor<fp16, [1, 1, 128, 64]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_268")];
            tensor<fp16, [1, 1, 128, 64]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 64, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 64, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 64, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_21")];
            tensor<fp16, [1, 1, 64, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 64]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_267")];
            tensor<bool, [1, 1, 2048, 64]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 64]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_266")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_140_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_140 = read_state(input = key_cache)[name = string("coreml_update_state_140")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_265")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_141_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_141 = read_state(input = value_cache)[name = string("coreml_update_state_141")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_140)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_141)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_698_0, tensor<fp16, [1, 8, 128, 64]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_264")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_263")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_140)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_142_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_142 = read_state(input = key_cache)[name = string("coreml_update_state_142")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_262")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_141)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_143_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_143 = read_state(input = value_cache)[name = string("coreml_update_state_143")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_142)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_143)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1019_0, tensor<fp16, [1, 8, 128, 64]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_261")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_260")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_142)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_144_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_144 = read_state(input = key_cache)[name = string("coreml_update_state_144")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_259")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_143)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_145_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_145 = read_state(input = value_cache)[name = string("coreml_update_state_145")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_144)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_145)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1340_0, tensor<fp16, [1, 8, 128, 64]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_258")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_257")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_144)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_146_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_146 = read_state(input = key_cache)[name = string("coreml_update_state_146")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_256")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_145)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_147_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_147 = read_state(input = value_cache)[name = string("coreml_update_state_147")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_146)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_147)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1661_0, tensor<fp16, [1, 8, 128, 64]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_255")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_254")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_146)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_148_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_148 = read_state(input = key_cache)[name = string("coreml_update_state_148")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_253")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_147)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_149_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_149 = read_state(input = value_cache)[name = string("coreml_update_state_149")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_148)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_149)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1982_0, tensor<fp16, [1, 8, 128, 64]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_252")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_251")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_148)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_150_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_150 = read_state(input = key_cache)[name = string("coreml_update_state_150")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_250")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_149)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_151_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_151 = read_state(input = value_cache)[name = string("coreml_update_state_151")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_150)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_151)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2303_0, tensor<fp16, [1, 8, 128, 64]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_249")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_248")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_150)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_152_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_152 = read_state(input = key_cache)[name = string("coreml_update_state_152")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_247")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_151)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_153_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_153 = read_state(input = value_cache)[name = string("coreml_update_state_153")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_152)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_153)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2624_0, tensor<fp16, [1, 8, 128, 64]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_246")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_245")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_152)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_154_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_154 = read_state(input = key_cache)[name = string("coreml_update_state_154")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_244")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_153)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_155_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_155 = read_state(input = value_cache)[name = string("coreml_update_state_155")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_154)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_155)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2945_0, tensor<fp16, [1, 8, 128, 64]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_243")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_242")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_154)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_156_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_156 = read_state(input = key_cache)[name = string("coreml_update_state_156")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_241")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_155)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_157_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_157 = read_state(input = value_cache)[name = string("coreml_update_state_157")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_156)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_157)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3266_0, tensor<fp16, [1, 8, 128, 64]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_240")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_156)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_158_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_158 = read_state(input = key_cache)[name = string("coreml_update_state_158")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_157)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_159_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_159 = read_state(input = value_cache)[name = string("coreml_update_state_159")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_158)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_159)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3587_0, tensor<fp16, [1, 8, 128, 64]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_158)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_160_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_160 = read_state(input = key_cache)[name = string("coreml_update_state_160")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_159)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_161_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_161 = read_state(input = value_cache)[name = string("coreml_update_state_161")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_160)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_161)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3908_0, tensor<fp16, [1, 8, 128, 64]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_160)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_162_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_162 = read_state(input = key_cache)[name = string("coreml_update_state_162")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_161)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_163_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_163 = read_state(input = value_cache)[name = string("coreml_update_state_163")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_162)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_163)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_4229_0, tensor<fp16, [1, 8, 128, 64]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_162)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_164_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_164 = read_state(input = key_cache)[name = string("coreml_update_state_164")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_163)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_165_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_165 = read_state(input = value_cache)[name = string("coreml_update_state_165")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_164)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_165)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_4550_0, tensor<fp16, [1, 8, 128, 64]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_164)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_166_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_166 = read_state(input = key_cache)[name = string("coreml_update_state_166")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_165)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_167_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_167 = read_state(input = value_cache)[name = string("coreml_update_state_167")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_166)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_167)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_4871_0, tensor<fp16, [1, 8, 128, 64]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states, tensor<fp16, [1, 2048, 1, 64]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
    func length_8<ios18>(tensor<fp32, [1, 2048, 1, 8]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [8]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(8)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [8]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 1, 128, 8]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 1, 128, 8]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 8, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 8, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 8, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_9")];
            tensor<fp16, [1, 1, 8, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 8]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_132")];
            tensor<bool, [1, 1, 2048, 8]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 8]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_56_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_57_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_56)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_57)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_698_0, tensor<fp16, [1, 8, 128, 8]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_56)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_58_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_57)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_59_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_58)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_59)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1019_0, tensor<fp16, [1, 8, 128, 8]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_58)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_59)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_60)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_61)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1340_0, tensor<fp16, [1, 8, 128, 8]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_60)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_62_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_61)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_63_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_62)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_63)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1661_0, tensor<fp16, [1, 8, 128, 8]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_62)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_64_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_64 = read_state(input = key_cache)[name = string("coreml_update_state_64")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_63)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_65_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_65 = read_state(input = value_cache)[name = string("coreml_update_state_65")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_64)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_65)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1982_0, tensor<fp16, [1, 8, 128, 8]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_64)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_66_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_66 = read_state(input = key_cache)[name = string("coreml_update_state_66")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_65)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_67_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_67 = read_state(input = value_cache)[name = string("coreml_update_state_67")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_66)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_67)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2303_0, tensor<fp16, [1, 8, 128, 8]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_66)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_68_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_68 = read_state(input = key_cache)[name = string("coreml_update_state_68")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_67)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_69_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_69 = read_state(input = value_cache)[name = string("coreml_update_state_69")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_68)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_69)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2624_0, tensor<fp16, [1, 8, 128, 8]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_68)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_70_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_70 = read_state(input = key_cache)[name = string("coreml_update_state_70")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_69)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_71_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_71 = read_state(input = value_cache)[name = string("coreml_update_state_71")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_70)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_71)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2945_0, tensor<fp16, [1, 8, 128, 8]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_70)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_72_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_72 = read_state(input = key_cache)[name = string("coreml_update_state_72")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_71)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_73_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_73 = read_state(input = value_cache)[name = string("coreml_update_state_73")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_72)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_73)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3266_0, tensor<fp16, [1, 8, 128, 8]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_72)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_74 = read_state(input = key_cache)[name = string("coreml_update_state_74")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_73)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_75 = read_state(input = value_cache)[name = string("coreml_update_state_75")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_74)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_75)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3587_0, tensor<fp16, [1, 8, 128, 8]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_74)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_76 = read_state(input = key_cache)[name = string("coreml_update_state_76")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_75)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_77 = read_state(input = value_cache)[name = string("coreml_update_state_77")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_76)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_77)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3908_0, tensor<fp16, [1, 8, 128, 8]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_76)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_78 = read_state(input = key_cache)[name = string("coreml_update_state_78")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_77)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_79 = read_state(input = value_cache)[name = string("coreml_update_state_79")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_78)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_79)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4229_0, tensor<fp16, [1, 8, 128, 8]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_78)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_80 = read_state(input = key_cache)[name = string("coreml_update_state_80")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_79)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_81 = read_state(input = value_cache)[name = string("coreml_update_state_81")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_80)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_81)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4550_0, tensor<fp16, [1, 8, 128, 8]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_80)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_82 = read_state(input = key_cache)[name = string("coreml_update_state_82")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_81)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_83 = read_state(input = value_cache)[name = string("coreml_update_state_83")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_82)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_83)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4871_0, tensor<fp16, [1, 8, 128, 8]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states, tensor<fp16, [1, 2048, 1, 8]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
    func main<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_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(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_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(4725952))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8924480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8920320))))[name = string("layers_0_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8926592))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21521920))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21509568))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21528128))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34123456))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34111104))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34129664))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50917440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50913280))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50919552))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51443904))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            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(51444800))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969728))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51969152))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51970048))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56168576))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56164416))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56170688))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68766016))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68753664))))[name = string("layers_1_mlp_gate_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(68772224))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81367552))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81355200))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81373760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93960896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93956736))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93963008))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98161536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157376))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98688000))))[name = string("layers_2_self_attn_k_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(98688896))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213824))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99213248))))[name = string("layers_2_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99214144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103412672))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408512))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103414784))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997760))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116004032))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120202560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120198400))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120204672))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729600))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120729024))))[name = string("layers_3_self_attn_k_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(120729920))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254848))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121254272))))[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(121255168))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125453696))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125449536))))[name = string("layers_3_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125455808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138051136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138038784))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138057344))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150652672))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150640320))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150658880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241856))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163248128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167446656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167442496))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167448768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167973120))))[name = string("layers_4_self_attn_k_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(167974016))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168498368))))[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(168499264))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172697792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172693632))))[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(172699904))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185295232))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185282880))))[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(185301440))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197896768))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197884416))))[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(197902976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210490112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210485952))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210492224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214690752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214686592))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214692864))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217792))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215217216))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215218112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227813440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227801088))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227819648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240414976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240402624))))[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(240421184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253008320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004160))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253010432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257208960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257204800))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257211072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257735424))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257736320))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261934848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261930688))))[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(261936960))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274532288))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274519936))))[name = string("layers_6_mlp_gate_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(274538496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287125632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287121472))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287127744))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291326272))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291322112))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291328384))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853312))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291852736))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291853632))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296052160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296048000))))[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(296054272))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308649600))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308637248))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_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(308655808))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321251136))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321238784))))[name = string("layers_7_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_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(321257344))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333844480))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333840320))))[name = string("layers_7_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333846592))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338045120))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040960))))[name = string("layers_8_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338047232))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338572160))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338571584))))[name = string("layers_8_self_attn_k_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(338572480))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351167808))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351155456))))[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(351174016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363769344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363756992))))[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(363775552))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376362688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376358528))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376364800))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380563328))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380559168))))[name = string("layers_9_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380565440))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381090368))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381089792))))[name = string("layers_9_self_attn_k_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(381090688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385289216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385285056))))[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(385291328))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397886656))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397874304))))[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(397892864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410488192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410475840))))[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(410494400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423081536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423077376))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423083648))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427282176))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427278016))))[name = string("layers_10_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427284288))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427808640))))[name = string("layers_10_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427809536))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432008064))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432003904))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432010176))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444605504))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444593152))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444611712))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457207040))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457194688))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457213248))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469800384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469796224))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469802496))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474001024))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473996864))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474003136))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528064))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474527488))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474528384))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478726912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478722752))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478729024))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491324352))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491312000))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491330560))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503925888))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503913536))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503932096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508130624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508126464))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508132736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657088))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508657984))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512856512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512852352))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512858624))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525453952))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525441600))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(1)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_425 = const()[name = string("op_425"), val = int32(0)];
            bool var_427_exclusive_0 = const()[name = string("op_427_exclusive_0"), val = bool(false)];
            bool var_427_reverse_0 = const()[name = string("op_427_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_427_cast_fp16 = cumsum(axis = var_425, exclusive = var_427_exclusive_0, reverse = var_427_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_427_cast_fp16")];
            fp16 var_429_promoted_to_fp16 = const()[name = string("op_429_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_427_cast_fp16, y = var_429_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_432_axes_0 = const()[name = string("op_432_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_432_cast_fp16 = expand_dims(axes = var_432_axes_0, x = position_offsets_cast_fp16)[name = string("op_432_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_432_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(525460160)))];
            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(533848832)))];
            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_451_perm_0 = const()[name = string("op_451_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_451_cast_fp16 = transpose(perm = var_451_perm_0, x = cos_1_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 1, 128, 1]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_451_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [3]> var_458_perm_0 = const()[name = string("op_458_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_460_axes_0 = const()[name = string("op_460_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_458_cast_fp16 = transpose(perm = var_458_perm_0, x = sin_1_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 1, 128, 1]> var_460_cast_fp16 = expand_dims(axes = var_460_axes_0, x = var_458_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_479 = expand_dims(axes = var_479_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_479")];
            tensor<int32, [1, 1, 2048]> var_472 = const()[name = string("op_472"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542237504)))];
            tensor<bool, [1, 1, 2048]> var_480 = greater(x = var_472, y = var_479)[name = string("op_480")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_480_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_480)[name = string("cast_1")];
            tensor<fp16, [1, 1, 1, 2048]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = var_480_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_491_promoted_to_fp16 = const()[name = string("op_491_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_487_cast_fp16)[name = string("transpose_42")];
            tensor<bool, [1, 1, 2048, 1]> var_492_cast_fp16 = equal(x = mask_cast_fp16, y = var_491_promoted_to_fp16)[name = string("op_492_cast_fp16")];
            fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_493_to_fp16, cond = var_492_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_503_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_503_cast_fp16")];
            int32 var_501 = const()[name = string("op_501"), 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_501, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_503_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(542245760)))];
            fp16 var_513_to_fp16 = const()[name = string("op_513_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_513_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_524_split_sizes_0 = const()[name = string("op_524_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_524_axis_0 = const()[name = string("op_524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_524_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_524_cast_fp16_1 = split(axis = var_524_axis_0, split_sizes = var_524_split_sizes_0, x = out_1_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_524_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(542254016)))];
            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_524_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            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_cast_fp16, x = var_524_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_581_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_581_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_588_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_592_cast_fp16 = mul(x = x_1_cast_fp16, y = var_453_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [2]> var_593_split_sizes_0 = const()[name = string("op_593_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_593_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_593_cast_fp16_1 = split(axis = var_593_axis_0, split_sizes = var_593_split_sizes_0, x = x_1_cast_fp16)[name = string("op_593_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_595_cast_fp16 = mul(x = var_593_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_595_cast_fp16")];
            int32 var_597 = const()[name = string("op_597"), val = int32(-2)];
            bool var_598_interleave_0 = const()[name = string("op_598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_598_cast_fp16 = concat(axis = var_597, interleave = var_598_interleave_0, values = (var_595_cast_fp16, var_593_cast_fp16_0))[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_599_cast_fp16 = mul(x = var_598_cast_fp16, y = var_460_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_592_cast_fp16, y = var_599_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_605_cast_fp16 = mul(x = var_581_cast_fp16, y = var_453_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<int32, [2]> var_606_split_sizes_0 = const()[name = string("op_606_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_606_axis_0 = const()[name = string("op_606_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_606_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_606_cast_fp16_1 = split(axis = var_606_axis_0, split_sizes = var_606_split_sizes_0, x = var_581_cast_fp16)[name = string("op_606_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_608_cast_fp16 = mul(x = var_606_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_608_cast_fp16")];
            int32 var_610 = const()[name = string("op_610"), val = int32(-2)];
            bool var_611_interleave_0 = const()[name = string("op_611_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_611_cast_fp16 = concat(axis = var_610, interleave = var_611_interleave_0, values = (var_608_cast_fp16, var_606_cast_fp16_0))[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_612_cast_fp16 = mul(x = var_611_cast_fp16, y = var_460_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_605_cast_fp16, y = var_612_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_0_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_0 = read_state(input = key_cache)[name = string("coreml_update_state_0")];
            tensor<fp16, [14, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_588_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_1_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_1 = read_state(input = value_cache)[name = string("coreml_update_state_1")];
            tensor<int32, [4]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = coreml_update_state_0)[name = string("op_682_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_685_axis_0 = const()[name = string("op_685_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_685_cast_fp16_1 = split(axis = var_685_axis_0, split_sizes = tile_0, x = var_682_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = coreml_update_state_1)[name = string("op_692_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_695_axis_0 = const()[name = string("op_695_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = tile_1, x = var_692_cast_fp16)[name = string("op_695_cast_fp16")];
            tensor<int32, [2]> var_698_split_sizes_0 = const()[name = string("op_698_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_698_axis_0 = const()[name = string("op_698_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_698_0, tensor<fp16, [1, 8, 128, 1]> var_698_1 = split(axis = var_698_axis_0, split_sizes = var_698_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_698")];
            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_685_cast_fp16_0, y = var_698_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_701_to_fp16 = const()[name = string("op_701_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_701_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_705 = const()[name = string("op_705"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_705, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_711_transpose_x_1 = const()[name = string("op_711_transpose_x_1"), val = bool(true)];
            bool var_711_transpose_y_1 = const()[name = string("op_711_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_711_cast_fp16 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_695_cast_fp16_0)[name = string("op_711_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_685_cast_fp16_1, y = var_698_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_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_713_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_717 = const()[name = string("op_717"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_717, 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_695_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_725 = const()[name = string("op_725"), 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_725, interleave = attn_output_3_interleave_0, values = (var_711_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_729_perm_0 = const()[name = string("op_729_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_729_cast_fp16 = transpose(perm = var_729_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_729_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            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_cast_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_762_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_762_cast_fp16")];
            int32 var_760 = const()[name = string("op_760"), 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_760, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_762_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(543302656)))];
            fp16 var_772_to_fp16 = const()[name = string("op_772_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_772_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_783_split_sizes_0 = const()[name = string("op_783_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_783_axis_0 = const()[name = string("op_783_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_783_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_783_cast_fp16_1 = split(axis = var_783_axis_0, split_sizes = var_783_split_sizes_0, x = out_3_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_800_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_800_cast_fp16")];
            tensor<int32, [2]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_806_cast_fp16 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_783_cast_fp16_0)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_800_cast_fp16, y = var_806_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_824_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_824_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), 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_822, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_824_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(543310912)))];
            fp16 var_834_to_fp16 = const()[name = string("op_834_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_834_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_845_split_sizes_0 = const()[name = string("op_845_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_845_axis_0 = const()[name = string("op_845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_845_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_845_cast_fp16_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = out_5_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_845_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_845_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_845_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_902_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_902_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_909_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_913_cast_fp16 = mul(x = x_11_cast_fp16, y = var_453_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [2]> var_914_split_sizes_0 = const()[name = string("op_914_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_914_axis_0 = const()[name = string("op_914_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_914_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_914_cast_fp16_1 = split(axis = var_914_axis_0, split_sizes = var_914_split_sizes_0, x = x_11_cast_fp16)[name = string("op_914_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_916_cast_fp16 = mul(x = var_914_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_916_cast_fp16")];
            int32 var_918 = const()[name = string("op_918"), val = int32(-2)];
            bool var_919_interleave_0 = const()[name = string("op_919_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_919_cast_fp16 = concat(axis = var_918, interleave = var_919_interleave_0, values = (var_916_cast_fp16, var_914_cast_fp16_0))[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_920_cast_fp16 = mul(x = var_919_cast_fp16, y = var_460_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_920_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_926_cast_fp16 = mul(x = var_902_cast_fp16, y = var_453_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [2]> var_927_split_sizes_0 = const()[name = string("op_927_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_927_axis_0 = const()[name = string("op_927_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_927_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_927_cast_fp16_1 = split(axis = var_927_axis_0, split_sizes = var_927_split_sizes_0, x = var_902_cast_fp16)[name = string("op_927_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_929_cast_fp16 = mul(x = var_927_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            int32 var_931 = const()[name = string("op_931"), val = int32(-2)];
            bool var_932_interleave_0 = const()[name = string("op_932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_932_cast_fp16 = concat(axis = var_931, interleave = var_932_interleave_0, values = (var_929_cast_fp16, var_927_cast_fp16_0))[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_933_cast_fp16 = mul(x = var_932_cast_fp16, y = var_460_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_926_cast_fp16, y = var_933_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_0)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_2_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_2 = read_state(input = key_cache)[name = string("coreml_update_state_2")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_909_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_1)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_3_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_3 = read_state(input = value_cache)[name = string("coreml_update_state_3")];
            tensor<int32, [4]> var_1003_begin_0 = const()[name = string("op_1003_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1003_end_0 = const()[name = string("op_1003_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1003_end_mask_0 = const()[name = string("op_1003_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = coreml_update_state_2)[name = string("op_1003_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_axis_0 = const()[name = string("op_1006_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1006_cast_fp16_1 = split(axis = var_1006_axis_0, split_sizes = tile_2, x = var_1003_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = coreml_update_state_3)[name = string("op_1013_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1016_axis_0 = const()[name = string("op_1016_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1016_cast_fp16_1 = split(axis = var_1016_axis_0, split_sizes = tile_3, x = var_1013_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<int32, [2]> var_1019_split_sizes_0 = const()[name = string("op_1019_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1019_axis_0 = const()[name = string("op_1019_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1019_0, tensor<fp16, [1, 8, 128, 1]> var_1019_1 = split(axis = var_1019_axis_0, split_sizes = var_1019_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1019")];
            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_1006_cast_fp16_0, y = var_1019_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1022_to_fp16 = const()[name = string("op_1022_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_1022_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_1026 = const()[name = string("op_1026"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_1026, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1032_transpose_x_1 = const()[name = string("op_1032_transpose_x_1"), val = bool(true)];
            bool var_1032_transpose_y_1 = const()[name = string("op_1032_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1032_cast_fp16 = matmul(transpose_x = var_1032_transpose_x_1, transpose_y = var_1032_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1016_cast_fp16_0)[name = string("op_1032_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_1006_cast_fp16_1, y = var_1019_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1034_to_fp16 = const()[name = string("op_1034_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_1034_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_1038 = const()[name = string("op_1038"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_1038, 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_1016_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1046 = const()[name = string("op_1046"), 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_1046, interleave = attn_output_11_interleave_0, values = (var_1032_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1050_perm_0 = const()[name = string("op_1050_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_1050_cast_fp16 = transpose(perm = var_1050_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1050_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1083_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1083_cast_fp16")];
            int32 var_1081 = const()[name = string("op_1081"), 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_1081, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1083_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(543319168)))];
            fp16 var_1093_to_fp16 = const()[name = string("op_1093_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_1093_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1104_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1104_cast_fp16_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1121_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [2]> var_1127_strides_0 = const()[name = string("op_1127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1127_pad_type_0 = const()[name = string("op_1127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1127_pad_0 = const()[name = string("op_1127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1127_dilations_0 = const()[name = string("op_1127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1127_groups_0 = const()[name = string("op_1127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1127_cast_fp16 = conv(dilations = var_1127_dilations_0, groups = var_1127_groups_0, pad = var_1127_pad_0, pad_type = var_1127_pad_type_0, strides = var_1127_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1104_cast_fp16_0)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1127_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1145_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1145_cast_fp16")];
            int32 var_1143 = const()[name = string("op_1143"), 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_1143, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1145_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(543327424)))];
            fp16 var_1155_to_fp16 = const()[name = string("op_1155_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_1155_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1166_split_sizes_0 = const()[name = string("op_1166_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1166_axis_0 = const()[name = string("op_1166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1166_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1166_cast_fp16_1 = split(axis = var_1166_axis_0, split_sizes = var_1166_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1166_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1166_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_1166_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_1223_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1223_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_1230_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1234_cast_fp16 = mul(x = x_21_cast_fp16, y = var_453_cast_fp16)[name = string("op_1234_cast_fp16")];
            tensor<int32, [2]> var_1235_split_sizes_0 = const()[name = string("op_1235_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1235_axis_0 = const()[name = string("op_1235_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1235_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1235_cast_fp16_1 = split(axis = var_1235_axis_0, split_sizes = var_1235_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1235_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_1237_cast_fp16 = mul(x = var_1235_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1237_cast_fp16")];
            int32 var_1239 = const()[name = string("op_1239"), val = int32(-2)];
            bool var_1240_interleave_0 = const()[name = string("op_1240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1240_cast_fp16 = concat(axis = var_1239, interleave = var_1240_interleave_0, values = (var_1237_cast_fp16, var_1235_cast_fp16_0))[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1241_cast_fp16 = mul(x = var_1240_cast_fp16, y = var_460_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1234_cast_fp16, y = var_1241_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1247_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_453_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [2]> var_1248_split_sizes_0 = const()[name = string("op_1248_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1248_axis_0 = const()[name = string("op_1248_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1248_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1248_cast_fp16_1 = split(axis = var_1248_axis_0, split_sizes = var_1248_split_sizes_0, x = var_1223_cast_fp16)[name = string("op_1248_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_1250_cast_fp16 = mul(x = var_1248_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1250_cast_fp16")];
            int32 var_1252 = const()[name = string("op_1252"), val = int32(-2)];
            bool var_1253_interleave_0 = const()[name = string("op_1253_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1253_cast_fp16 = concat(axis = var_1252, interleave = var_1253_interleave_0, values = (var_1250_cast_fp16, var_1248_cast_fp16_0))[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1254_cast_fp16 = mul(x = var_1253_cast_fp16, y = var_460_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1254_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_2)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_4_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_4 = read_state(input = key_cache)[name = string("coreml_update_state_4")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1230_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_3)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_5_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_5 = read_state(input = value_cache)[name = string("coreml_update_state_5")];
            tensor<int32, [4]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = coreml_update_state_4)[name = string("op_1324_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1327_axis_0 = const()[name = string("op_1327_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1327_cast_fp16_1 = split(axis = var_1327_axis_0, split_sizes = tile_4, x = var_1324_cast_fp16)[name = string("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1334_begin_0 = const()[name = string("op_1334_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1334_end_0 = const()[name = string("op_1334_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1334_end_mask_0 = const()[name = string("op_1334_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1334_cast_fp16 = slice_by_index(begin = var_1334_begin_0, end = var_1334_end_0, end_mask = var_1334_end_mask_0, x = coreml_update_state_5)[name = string("op_1334_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1337_axis_0 = const()[name = string("op_1337_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1337_cast_fp16_1 = split(axis = var_1337_axis_0, split_sizes = tile_5, x = var_1334_cast_fp16)[name = string("op_1337_cast_fp16")];
            tensor<int32, [2]> var_1340_split_sizes_0 = const()[name = string("op_1340_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1340_axis_0 = const()[name = string("op_1340_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1340_0, tensor<fp16, [1, 8, 128, 1]> var_1340_1 = split(axis = var_1340_axis_0, split_sizes = var_1340_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1340")];
            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_1327_cast_fp16_0, y = var_1340_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1343_to_fp16 = const()[name = string("op_1343_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_1343_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_1347 = const()[name = string("op_1347"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1347, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1353_transpose_x_1 = const()[name = string("op_1353_transpose_x_1"), val = bool(true)];
            bool var_1353_transpose_y_1 = const()[name = string("op_1353_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1353_cast_fp16 = matmul(transpose_x = var_1353_transpose_x_1, transpose_y = var_1353_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1337_cast_fp16_0)[name = string("op_1353_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_1327_cast_fp16_1, y = var_1340_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1355_to_fp16 = const()[name = string("op_1355_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_1355_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_1359 = const()[name = string("op_1359"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1359, 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_1337_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1367 = const()[name = string("op_1367"), 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_1367, interleave = attn_output_19_interleave_0, values = (var_1353_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1371_perm_0 = const()[name = string("op_1371_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_1371_cast_fp16 = transpose(perm = var_1371_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1371_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1404_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1404_cast_fp16")];
            int32 var_1402 = const()[name = string("op_1402"), 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_1402, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1404_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(543335680)))];
            fp16 var_1414_to_fp16 = const()[name = string("op_1414_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_1414_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1425_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1425_cast_fp16_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1425_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(543343936)))];
            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_1425_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1442_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568509824)))];
            tensor<int32, [2]> var_1448_strides_0 = const()[name = string("op_1448_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1448_pad_type_0 = const()[name = string("op_1448_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1448_pad_0 = const()[name = string("op_1448_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1448_dilations_0 = const()[name = string("op_1448_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1448_groups_0 = const()[name = string("op_1448_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1448_cast_fp16 = conv(dilations = var_1448_dilations_0, groups = var_1448_groups_0, pad = var_1448_pad_0, pad_type = var_1448_pad_type_0, strides = var_1448_strides_0, weight = layers_2_mlp_up_proj_weight_to_fp16, x = var_1425_cast_fp16_0)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1442_cast_fp16, y = var_1448_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1466_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1466_cast_fp16")];
            int32 var_1464 = const()[name = string("op_1464"), 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_1464, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1466_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(593675712)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_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_1476_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1487_split_sizes_0 = const()[name = string("op_1487_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1487_axis_0 = const()[name = string("op_1487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1487_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1487_cast_fp16_1 = split(axis = var_1487_axis_0, split_sizes = var_1487_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1487_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1487_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1487_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_1487_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_1544_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1544_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_1551_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1555_cast_fp16 = mul(x = x_31_cast_fp16, y = var_453_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<int32, [2]> var_1556_split_sizes_0 = const()[name = string("op_1556_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1556_axis_0 = const()[name = string("op_1556_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1556_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1556_cast_fp16_1 = split(axis = var_1556_axis_0, split_sizes = var_1556_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1556_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_1558_cast_fp16 = mul(x = var_1556_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1558_cast_fp16")];
            int32 var_1560 = const()[name = string("op_1560"), val = int32(-2)];
            bool var_1561_interleave_0 = const()[name = string("op_1561_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1561_cast_fp16 = concat(axis = var_1560, interleave = var_1561_interleave_0, values = (var_1558_cast_fp16, var_1556_cast_fp16_0))[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1562_cast_fp16 = mul(x = var_1561_cast_fp16, y = var_460_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1555_cast_fp16, y = var_1562_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1568_cast_fp16 = mul(x = var_1544_cast_fp16, y = var_453_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<int32, [2]> var_1569_split_sizes_0 = const()[name = string("op_1569_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1569_axis_0 = const()[name = string("op_1569_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1569_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1569_cast_fp16_1 = split(axis = var_1569_axis_0, split_sizes = var_1569_split_sizes_0, x = var_1544_cast_fp16)[name = string("op_1569_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_1571_cast_fp16 = mul(x = var_1569_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1571_cast_fp16")];
            int32 var_1573 = const()[name = string("op_1573"), val = int32(-2)];
            bool var_1574_interleave_0 = const()[name = string("op_1574_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1571_cast_fp16, var_1569_cast_fp16_0))[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1575_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_460_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1568_cast_fp16, y = var_1575_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_4)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_6_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_6 = read_state(input = key_cache)[name = string("coreml_update_state_6")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1551_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_5)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_7_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_7 = read_state(input = value_cache)[name = string("coreml_update_state_7")];
            tensor<int32, [4]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = coreml_update_state_6)[name = string("op_1645_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_axis_0 = const()[name = string("op_1648_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1648_cast_fp16_1 = split(axis = var_1648_axis_0, split_sizes = tile_6, x = var_1645_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [4]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = coreml_update_state_7)[name = string("op_1655_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1658_axis_0 = const()[name = string("op_1658_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1658_cast_fp16_1 = split(axis = var_1658_axis_0, split_sizes = tile_7, x = var_1655_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [2]> var_1661_split_sizes_0 = const()[name = string("op_1661_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1661_axis_0 = const()[name = string("op_1661_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1661_0, tensor<fp16, [1, 8, 128, 1]> var_1661_1 = split(axis = var_1661_axis_0, split_sizes = var_1661_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1661")];
            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_1648_cast_fp16_0, y = var_1661_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_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_1664_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_1668 = const()[name = string("op_1668"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1668, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1674_transpose_x_1 = const()[name = string("op_1674_transpose_x_1"), val = bool(true)];
            bool var_1674_transpose_y_1 = const()[name = string("op_1674_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1674_cast_fp16 = matmul(transpose_x = var_1674_transpose_x_1, transpose_y = var_1674_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1658_cast_fp16_0)[name = string("op_1674_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_1648_cast_fp16_1, y = var_1661_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1676_to_fp16 = const()[name = string("op_1676_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_1676_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_1680 = const()[name = string("op_1680"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1680, 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_1658_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), 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_1688, interleave = attn_output_27_interleave_0, values = (var_1674_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1692_perm_0 = const()[name = string("op_1692_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_1692_cast_fp16 = transpose(perm = var_1692_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1692_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_1725_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1725_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), 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_1723, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1725_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(593683968)))];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_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_1735_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1746_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1746_cast_fp16_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1763_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [2]> var_1769_strides_0 = const()[name = string("op_1769_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1769_pad_type_0 = const()[name = string("op_1769_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1769_pad_0 = const()[name = string("op_1769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1769_dilations_0 = const()[name = string("op_1769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1769_groups_0 = const()[name = string("op_1769_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1769_cast_fp16 = conv(dilations = var_1769_dilations_0, groups = var_1769_groups_0, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1769_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1746_cast_fp16_0)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1763_cast_fp16, y = var_1769_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_1787_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1787_cast_fp16")];
            int32 var_1785 = const()[name = string("op_1785"), 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_1785, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1787_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(593692224)))];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_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_1797_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1808_split_sizes_0 = const()[name = string("op_1808_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1808_axis_0 = const()[name = string("op_1808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1808_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1808_cast_fp16_1 = split(axis = var_1808_axis_0, split_sizes = var_1808_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1808_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1808_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_1808_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_1865_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1865_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_1872_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1876_cast_fp16 = mul(x = x_41_cast_fp16, y = var_453_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [2]> var_1877_split_sizes_0 = const()[name = string("op_1877_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1877_axis_0 = const()[name = string("op_1877_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1877_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1877_cast_fp16_1 = split(axis = var_1877_axis_0, split_sizes = var_1877_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1877_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_1879_cast_fp16 = mul(x = var_1877_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1879_cast_fp16")];
            int32 var_1881 = const()[name = string("op_1881"), val = int32(-2)];
            bool var_1882_interleave_0 = const()[name = string("op_1882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1882_cast_fp16 = concat(axis = var_1881, interleave = var_1882_interleave_0, values = (var_1879_cast_fp16, var_1877_cast_fp16_0))[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1883_cast_fp16 = mul(x = var_1882_cast_fp16, y = var_460_cast_fp16)[name = string("op_1883_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1876_cast_fp16, y = var_1883_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1889_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_453_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [2]> var_1890_split_sizes_0 = const()[name = string("op_1890_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1890_axis_0 = const()[name = string("op_1890_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1890_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1890_cast_fp16_1 = split(axis = var_1890_axis_0, split_sizes = var_1890_split_sizes_0, x = var_1865_cast_fp16)[name = string("op_1890_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_1892_cast_fp16 = mul(x = var_1890_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            int32 var_1894 = const()[name = string("op_1894"), val = int32(-2)];
            bool var_1895_interleave_0 = const()[name = string("op_1895_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1895_cast_fp16 = concat(axis = var_1894, interleave = var_1895_interleave_0, values = (var_1892_cast_fp16, var_1890_cast_fp16_0))[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1896_cast_fp16 = mul(x = var_1895_cast_fp16, y = var_460_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1896_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_6)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_8_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_8 = read_state(input = key_cache)[name = string("coreml_update_state_8")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1872_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_7)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_9_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_9 = read_state(input = value_cache)[name = string("coreml_update_state_9")];
            tensor<int32, [4]> var_1966_begin_0 = const()[name = string("op_1966_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1966_end_0 = const()[name = string("op_1966_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = string("op_1966_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = coreml_update_state_8)[name = string("op_1966_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1969_axis_0 = const()[name = string("op_1969_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1969_cast_fp16_1 = split(axis = var_1969_axis_0, split_sizes = tile_8, x = var_1966_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<int32, [4]> var_1976_begin_0 = const()[name = string("op_1976_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1976_end_0 = const()[name = string("op_1976_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1976_end_mask_0 = const()[name = string("op_1976_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1976_cast_fp16 = slice_by_index(begin = var_1976_begin_0, end = var_1976_end_0, end_mask = var_1976_end_mask_0, x = coreml_update_state_9)[name = string("op_1976_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1979_axis_0 = const()[name = string("op_1979_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1979_cast_fp16_1 = split(axis = var_1979_axis_0, split_sizes = tile_9, x = var_1976_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<int32, [2]> var_1982_split_sizes_0 = const()[name = string("op_1982_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1982_axis_0 = const()[name = string("op_1982_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1982_0, tensor<fp16, [1, 8, 128, 1]> var_1982_1 = split(axis = var_1982_axis_0, split_sizes = var_1982_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1982")];
            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_1969_cast_fp16_0, y = var_1982_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_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_1985_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_1989 = const()[name = string("op_1989"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1989, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1995_transpose_x_1 = const()[name = string("op_1995_transpose_x_1"), val = bool(true)];
            bool var_1995_transpose_y_1 = const()[name = string("op_1995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1995_cast_fp16 = matmul(transpose_x = var_1995_transpose_x_1, transpose_y = var_1995_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1979_cast_fp16_0)[name = string("op_1995_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_1969_cast_fp16_1, y = var_1982_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1997_to_fp16 = const()[name = string("op_1997_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_1997_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_2001 = const()[name = string("op_2001"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_2001, 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_1979_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2009 = const()[name = string("op_2009"), 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_2009, interleave = attn_output_35_interleave_0, values = (var_1995_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2013_perm_0 = const()[name = string("op_2013_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_2013_cast_fp16 = transpose(perm = var_2013_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2013_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_2046_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2044 = const()[name = string("op_2044"), 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_2044, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2046_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(593700480)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_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_2056_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2067_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2067_cast_fp16_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2067_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_2067_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2084_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [2]> var_2090_strides_0 = const()[name = string("op_2090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2090_pad_type_0 = const()[name = string("op_2090_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2090_pad_0 = const()[name = string("op_2090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2090_dilations_0 = const()[name = string("op_2090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2090_groups_0 = const()[name = string("op_2090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2090_cast_fp16 = conv(dilations = var_2090_dilations_0, groups = var_2090_groups_0, pad = var_2090_pad_0, pad_type = var_2090_pad_type_0, strides = var_2090_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2067_cast_fp16_0)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2090_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_2108_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2108_cast_fp16")];
            int32 var_2106 = const()[name = string("op_2106"), 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_2106, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2108_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(593708736)))];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_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_2118_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2129_split_sizes_0 = const()[name = string("op_2129_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2129_axis_0 = const()[name = string("op_2129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2129_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2129_cast_fp16_1 = split(axis = var_2129_axis_0, split_sizes = var_2129_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_2129_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593716992)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_2129_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_2186_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2186_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_2193_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2193_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2197_cast_fp16 = mul(x = x_51_cast_fp16, y = var_453_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [2]> var_2198_split_sizes_0 = const()[name = string("op_2198_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2198_axis_0 = const()[name = string("op_2198_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2198_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2198_cast_fp16_1 = split(axis = var_2198_axis_0, split_sizes = var_2198_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2198_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_2200_cast_fp16 = mul(x = var_2198_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            int32 var_2202 = const()[name = string("op_2202"), val = int32(-2)];
            bool var_2203_interleave_0 = const()[name = string("op_2203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2203_cast_fp16 = concat(axis = var_2202, interleave = var_2203_interleave_0, values = (var_2200_cast_fp16, var_2198_cast_fp16_0))[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2204_cast_fp16 = mul(x = var_2203_cast_fp16, y = var_460_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2197_cast_fp16, y = var_2204_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2210_cast_fp16 = mul(x = var_2186_cast_fp16, y = var_453_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [2]> var_2211_split_sizes_0 = const()[name = string("op_2211_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2211_axis_0 = const()[name = string("op_2211_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2211_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2211_cast_fp16_1 = split(axis = var_2211_axis_0, split_sizes = var_2211_split_sizes_0, x = var_2186_cast_fp16)[name = string("op_2211_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_2213_cast_fp16 = mul(x = var_2211_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2213_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(-2)];
            bool var_2216_interleave_0 = const()[name = string("op_2216_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2216_cast_fp16 = concat(axis = var_2215, interleave = var_2216_interleave_0, values = (var_2213_cast_fp16, var_2211_cast_fp16_0))[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2217_cast_fp16 = mul(x = var_2216_cast_fp16, y = var_460_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2210_cast_fp16, y = var_2217_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_8)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_10_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_10 = read_state(input = key_cache)[name = string("coreml_update_state_10")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2193_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_9)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_11_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_11 = read_state(input = value_cache)[name = string("coreml_update_state_11")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = coreml_update_state_10)[name = string("op_2287_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_axis_0 = const()[name = string("op_2290_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2290_cast_fp16_1 = split(axis = var_2290_axis_0, split_sizes = tile_10, x = var_2287_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = coreml_update_state_11)[name = string("op_2297_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2300_axis_0 = const()[name = string("op_2300_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2300_cast_fp16_1 = split(axis = var_2300_axis_0, split_sizes = tile_11, x = var_2297_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [2]> var_2303_split_sizes_0 = const()[name = string("op_2303_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2303_axis_0 = const()[name = string("op_2303_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2303_0, tensor<fp16, [1, 8, 128, 1]> var_2303_1 = split(axis = var_2303_axis_0, split_sizes = var_2303_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2303")];
            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_2290_cast_fp16_0, y = var_2303_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2306_to_fp16 = const()[name = string("op_2306_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_2306_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_2310 = const()[name = string("op_2310"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2310, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2316_transpose_x_1 = const()[name = string("op_2316_transpose_x_1"), val = bool(true)];
            bool var_2316_transpose_y_1 = const()[name = string("op_2316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2316_cast_fp16 = matmul(transpose_x = var_2316_transpose_x_1, transpose_y = var_2316_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2300_cast_fp16_0)[name = string("op_2316_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_2290_cast_fp16_1, y = var_2303_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2318_to_fp16 = const()[name = string("op_2318_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_2318_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_2322 = const()[name = string("op_2322"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2322, 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_2300_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2330 = const()[name = string("op_2330"), 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_2330, interleave = attn_output_43_interleave_0, values = (var_2316_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2334_perm_0 = const()[name = string("op_2334_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_2334_cast_fp16 = transpose(perm = var_2334_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2334_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594765632)))];
            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_to_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_2367_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2367_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), 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_2365, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2367_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(603154304)))];
            fp16 var_2377_to_fp16 = const()[name = string("op_2377_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_2377_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2388_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2388_cast_fp16_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2405_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [2]> var_2411_strides_0 = const()[name = string("op_2411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2411_pad_type_0 = const()[name = string("op_2411_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2411_pad_0 = const()[name = string("op_2411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2411_dilations_0 = const()[name = string("op_2411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2411_groups_0 = const()[name = string("op_2411_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2411_cast_fp16 = conv(dilations = var_2411_dilations_0, groups = var_2411_groups_0, pad = var_2411_pad_0, pad_type = var_2411_pad_type_0, strides = var_2411_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2388_cast_fp16_0)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2405_cast_fp16, y = var_2411_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_2429_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2429_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), 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_2427, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2429_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(603162560)))];
            fp16 var_2439_to_fp16 = const()[name = string("op_2439_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_2439_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2450_split_sizes_0 = const()[name = string("op_2450_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2450_axis_0 = const()[name = string("op_2450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2450_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2450_cast_fp16_1 = split(axis = var_2450_axis_0, split_sizes = var_2450_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2450_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603170816)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2450_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_2507_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2507_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_2514_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2518_cast_fp16 = mul(x = x_61_cast_fp16, y = var_453_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<int32, [2]> var_2519_split_sizes_0 = const()[name = string("op_2519_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2519_axis_0 = const()[name = string("op_2519_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2519_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2519_cast_fp16_1 = split(axis = var_2519_axis_0, split_sizes = var_2519_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2519_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_2521_cast_fp16 = mul(x = var_2519_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            int32 var_2523 = const()[name = string("op_2523"), val = int32(-2)];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2524_cast_fp16 = concat(axis = var_2523, interleave = var_2524_interleave_0, values = (var_2521_cast_fp16, var_2519_cast_fp16_0))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2525_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_460_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2525_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2531_cast_fp16 = mul(x = var_2507_cast_fp16, y = var_453_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [2]> var_2532_split_sizes_0 = const()[name = string("op_2532_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2532_axis_0 = const()[name = string("op_2532_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2532_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2532_cast_fp16_1 = split(axis = var_2532_axis_0, split_sizes = var_2532_split_sizes_0, x = var_2507_cast_fp16)[name = string("op_2532_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_2534_cast_fp16 = mul(x = var_2532_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2534_cast_fp16")];
            int32 var_2536 = const()[name = string("op_2536"), val = int32(-2)];
            bool var_2537_interleave_0 = const()[name = string("op_2537_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2537_cast_fp16 = concat(axis = var_2536, interleave = var_2537_interleave_0, values = (var_2534_cast_fp16, var_2532_cast_fp16_0))[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2538_cast_fp16 = mul(x = var_2537_cast_fp16, y = var_460_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2538_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_10)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_12_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_12 = read_state(input = key_cache)[name = string("coreml_update_state_12")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2514_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_11)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_13_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_13 = read_state(input = value_cache)[name = string("coreml_update_state_13")];
            tensor<int32, [4]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = coreml_update_state_12)[name = string("op_2608_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2611_axis_0 = const()[name = string("op_2611_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2611_cast_fp16_1 = split(axis = var_2611_axis_0, split_sizes = tile_12, x = var_2608_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [4]> var_2618_begin_0 = const()[name = string("op_2618_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2618_end_0 = const()[name = string("op_2618_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2618_end_mask_0 = const()[name = string("op_2618_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = coreml_update_state_13)[name = string("op_2618_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2621_axis_0 = const()[name = string("op_2621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2621_cast_fp16_1 = split(axis = var_2621_axis_0, split_sizes = tile_13, x = var_2618_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2624_0, tensor<fp16, [1, 8, 128, 1]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2624")];
            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_2611_cast_fp16_0, y = var_2624_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_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_2627_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_2631 = const()[name = string("op_2631"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2631, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2637_transpose_x_1 = const()[name = string("op_2637_transpose_x_1"), val = bool(true)];
            bool var_2637_transpose_y_1 = const()[name = string("op_2637_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2637_cast_fp16 = matmul(transpose_x = var_2637_transpose_x_1, transpose_y = var_2637_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2621_cast_fp16_0)[name = string("op_2637_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_2611_cast_fp16_1, y = var_2624_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2639_to_fp16 = const()[name = string("op_2639_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_2639_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_2643 = const()[name = string("op_2643"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2643, 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_2621_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2651 = const()[name = string("op_2651"), 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_2651, interleave = attn_output_51_interleave_0, values = (var_2637_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2655_perm_0 = const()[name = string("op_2655_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_2655_cast_fp16 = transpose(perm = var_2655_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2655_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_2688_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), 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_2686, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2688_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(604219456)))];
            fp16 var_2698_to_fp16 = const()[name = string("op_2698_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_2698_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2709_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2709_cast_fp16_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2709_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_2709_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2726_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604227712)))];
            tensor<int32, [2]> var_2732_strides_0 = const()[name = string("op_2732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2732_pad_type_0 = const()[name = string("op_2732_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2732_pad_0 = const()[name = string("op_2732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2732_dilations_0 = const()[name = string("op_2732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2732_groups_0 = const()[name = string("op_2732_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2732_cast_fp16 = conv(dilations = var_2732_dilations_0, groups = var_2732_groups_0, pad = var_2732_pad_0, pad_type = var_2732_pad_type_0, strides = var_2732_strides_0, weight = layers_6_mlp_up_proj_weight_to_fp16, x = var_2709_cast_fp16_0)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2732_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_2750_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2750_cast_fp16")];
            int32 var_2748 = const()[name = string("op_2748"), 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_2748, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2750_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(629393600)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_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_2760_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2771_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2771_cast_fp16_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2771_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2771_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629401856)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2771_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_2828_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2828_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_2835_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2839_cast_fp16 = mul(x = x_71_cast_fp16, y = var_453_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<int32, [2]> var_2840_split_sizes_0 = const()[name = string("op_2840_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2840_axis_0 = const()[name = string("op_2840_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2840_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2840_cast_fp16_1 = split(axis = var_2840_axis_0, split_sizes = var_2840_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2840_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_2842_cast_fp16 = mul(x = var_2840_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            int32 var_2844 = const()[name = string("op_2844"), val = int32(-2)];
            bool var_2845_interleave_0 = const()[name = string("op_2845_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2845_cast_fp16 = concat(axis = var_2844, interleave = var_2845_interleave_0, values = (var_2842_cast_fp16, var_2840_cast_fp16_0))[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2846_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_460_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2846_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2852_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_453_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = string("op_2853_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2853_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = var_2828_cast_fp16)[name = string("op_2853_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_2855_cast_fp16 = mul(x = var_2853_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            int32 var_2857 = const()[name = string("op_2857"), val = int32(-2)];
            bool var_2858_interleave_0 = const()[name = string("op_2858_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2858_cast_fp16 = concat(axis = var_2857, interleave = var_2858_interleave_0, values = (var_2855_cast_fp16, var_2853_cast_fp16_0))[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2859_cast_fp16 = mul(x = var_2858_cast_fp16, y = var_460_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2859_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_12)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_14_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_14 = read_state(input = key_cache)[name = string("coreml_update_state_14")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2835_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_13)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_15_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_15 = read_state(input = value_cache)[name = string("coreml_update_state_15")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = coreml_update_state_14)[name = string("op_2929_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2932_cast_fp16_1 = split(axis = var_2932_axis_0, split_sizes = tile_14, x = var_2929_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = coreml_update_state_15)[name = string("op_2939_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2942_axis_0 = const()[name = string("op_2942_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2942_cast_fp16_1 = split(axis = var_2942_axis_0, split_sizes = tile_15, x = var_2939_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [2]> var_2945_split_sizes_0 = const()[name = string("op_2945_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2945_axis_0 = const()[name = string("op_2945_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2945_0, tensor<fp16, [1, 8, 128, 1]> var_2945_1 = split(axis = var_2945_axis_0, split_sizes = var_2945_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2945")];
            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_2932_cast_fp16_0, y = var_2945_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2948_to_fp16 = const()[name = string("op_2948_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_2948_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_2952 = const()[name = string("op_2952"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2952, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2958_transpose_x_1 = const()[name = string("op_2958_transpose_x_1"), val = bool(true)];
            bool var_2958_transpose_y_1 = const()[name = string("op_2958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2958_cast_fp16 = matmul(transpose_x = var_2958_transpose_x_1, transpose_y = var_2958_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2942_cast_fp16_0)[name = string("op_2958_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_2932_cast_fp16_1, y = var_2945_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_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_2960_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_2964 = const()[name = string("op_2964"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2964, 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_2942_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2972 = const()[name = string("op_2972"), 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_2972, interleave = attn_output_59_interleave_0, values = (var_2958_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2976_perm_0 = const()[name = string("op_2976_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_2976_cast_fp16 = transpose(perm = var_2976_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2976_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_3009_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3009_cast_fp16")];
            int32 var_3007 = const()[name = string("op_3007"), 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_3007, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3009_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(630450496)))];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_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_3019_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3030_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3030_cast_fp16_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3030_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_3030_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3047_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [2]> var_3053_strides_0 = const()[name = string("op_3053_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3053_pad_type_0 = const()[name = string("op_3053_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3053_pad_0 = const()[name = string("op_3053_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3053_dilations_0 = const()[name = string("op_3053_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3053_groups_0 = const()[name = string("op_3053_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3053_cast_fp16 = conv(dilations = var_3053_dilations_0, groups = var_3053_groups_0, pad = var_3053_pad_0, pad_type = var_3053_pad_type_0, strides = var_3053_strides_0, weight = layers_7_mlp_up_proj_weight_cast_fp16, x = var_3030_cast_fp16_0)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_3047_cast_fp16, y = var_3053_cast_fp16)[name = string("x_79_cast_fp16")];
            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_cast_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_3071_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            int32 var_3069 = const()[name = string("op_3069"), 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_3069, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3071_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(630458752)))];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_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_3081_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3092_split_sizes_0 = const()[name = string("op_3092_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3092_axis_0 = const()[name = string("op_3092_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3092_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3092_cast_fp16_1 = split(axis = var_3092_axis_0, split_sizes = var_3092_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3092_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            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_cast_fp16, x = var_3092_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467008)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_3092_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_3149_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3149_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_3156_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3160_cast_fp16 = mul(x = x_81_cast_fp16, y = var_453_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [2]> var_3161_split_sizes_0 = const()[name = string("op_3161_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3161_axis_0 = const()[name = string("op_3161_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3161_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3161_cast_fp16_1 = split(axis = var_3161_axis_0, split_sizes = var_3161_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3161_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_3163_cast_fp16 = mul(x = var_3161_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
            int32 var_3165 = const()[name = string("op_3165"), val = int32(-2)];
            bool var_3166_interleave_0 = const()[name = string("op_3166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3166_cast_fp16 = concat(axis = var_3165, interleave = var_3166_interleave_0, values = (var_3163_cast_fp16, var_3161_cast_fp16_0))[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3167_cast_fp16 = mul(x = var_3166_cast_fp16, y = var_460_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3160_cast_fp16, y = var_3167_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3173_cast_fp16 = mul(x = var_3149_cast_fp16, y = var_453_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [2]> var_3174_split_sizes_0 = const()[name = string("op_3174_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3174_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3174_cast_fp16_1 = split(axis = var_3174_axis_0, split_sizes = var_3174_split_sizes_0, x = var_3149_cast_fp16)[name = string("op_3174_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_3176_cast_fp16 = mul(x = var_3174_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3176_cast_fp16")];
            int32 var_3178 = const()[name = string("op_3178"), val = int32(-2)];
            bool var_3179_interleave_0 = const()[name = string("op_3179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3179_cast_fp16 = concat(axis = var_3178, interleave = var_3179_interleave_0, values = (var_3176_cast_fp16, var_3174_cast_fp16_0))[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3180_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_460_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3180_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_14)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_16_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_16 = read_state(input = key_cache)[name = string("coreml_update_state_16")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3156_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_15)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_17_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_17 = read_state(input = value_cache)[name = string("coreml_update_state_17")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = string("op_3250_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = string("op_3250_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = string("op_3250_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = coreml_update_state_16)[name = string("op_3250_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3253_axis_0 = const()[name = string("op_3253_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3253_cast_fp16_1 = split(axis = var_3253_axis_0, split_sizes = tile_16, x = var_3250_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3260_begin_0 = const()[name = string("op_3260_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3260_end_0 = const()[name = string("op_3260_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3260_end_mask_0 = const()[name = string("op_3260_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3260_cast_fp16 = slice_by_index(begin = var_3260_begin_0, end = var_3260_end_0, end_mask = var_3260_end_mask_0, x = coreml_update_state_17)[name = string("op_3260_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3263_axis_0 = const()[name = string("op_3263_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3263_cast_fp16_1 = split(axis = var_3263_axis_0, split_sizes = tile_17, x = var_3260_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = string("op_3266_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3266_axis_0 = const()[name = string("op_3266_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3266_0, tensor<fp16, [1, 8, 128, 1]> var_3266_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3266")];
            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_3253_cast_fp16_0, y = var_3266_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3269_to_fp16 = const()[name = string("op_3269_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_3269_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_3273 = const()[name = string("op_3273"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3273, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3279_transpose_x_1 = const()[name = string("op_3279_transpose_x_1"), val = bool(true)];
            bool var_3279_transpose_y_1 = const()[name = string("op_3279_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3279_cast_fp16 = matmul(transpose_x = var_3279_transpose_x_1, transpose_y = var_3279_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3263_cast_fp16_0)[name = string("op_3279_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_3253_cast_fp16_1, y = var_3266_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_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_3281_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_3285 = const()[name = string("op_3285"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_143_cast_fp16 = softmax(axis = var_3285, 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_3263_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3293 = const()[name = string("op_3293"), 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_3293, interleave = attn_output_67_interleave_0, values = (var_3279_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3297_perm_0 = const()[name = string("op_3297_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_3297_cast_fp16 = transpose(perm = var_3297_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3297_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631515648)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_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_3330_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3330_cast_fp16")];
            int32 var_3328 = const()[name = string("op_3328"), 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_3328, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3330_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(639904320)))];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_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_3340_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3351_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3351_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_3351_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3368_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [2]> var_3374_strides_0 = const()[name = string("op_3374_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3374_pad_type_0 = const()[name = string("op_3374_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3374_pad_0 = const()[name = string("op_3374_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3374_dilations_0 = const()[name = string("op_3374_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3374_groups_0 = const()[name = string("op_3374_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3374_cast_fp16 = conv(dilations = var_3374_dilations_0, groups = var_3374_groups_0, pad = var_3374_pad_0, pad_type = var_3374_pad_type_0, strides = var_3374_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3351_cast_fp16_0)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_89_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3374_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_3392_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3392_cast_fp16")];
            int32 var_3390 = const()[name = string("op_3390"), 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_3390, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3392_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(639912576)))];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_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_3402_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3413_split_sizes_0 = const()[name = string("op_3413_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3413_axis_0 = const()[name = string("op_3413_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3413_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3413_cast_fp16_1 = split(axis = var_3413_axis_0, split_sizes = var_3413_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3413_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            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_cast_fp16, x = var_3413_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639920832)))];
            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_to_fp16, x = var_3413_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_3470_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3470_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_3477_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3481_cast_fp16 = mul(x = x_91_cast_fp16, y = var_453_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [2]> var_3482_split_sizes_0 = const()[name = string("op_3482_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3482_axis_0 = const()[name = string("op_3482_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3482_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3482_cast_fp16_1 = split(axis = var_3482_axis_0, split_sizes = var_3482_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3482_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_3484_cast_fp16 = mul(x = var_3482_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3484_cast_fp16")];
            int32 var_3486 = const()[name = string("op_3486"), val = int32(-2)];
            bool var_3487_interleave_0 = const()[name = string("op_3487_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3487_cast_fp16 = concat(axis = var_3486, interleave = var_3487_interleave_0, values = (var_3484_cast_fp16, var_3482_cast_fp16_0))[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3488_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_460_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_57_cast_fp16 = add(x = var_3481_cast_fp16, y = var_3488_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3494_cast_fp16 = mul(x = var_3470_cast_fp16, y = var_453_cast_fp16)[name = string("op_3494_cast_fp16")];
            tensor<int32, [2]> var_3495_split_sizes_0 = const()[name = string("op_3495_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3495_axis_0 = const()[name = string("op_3495_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3495_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3495_cast_fp16_1 = split(axis = var_3495_axis_0, split_sizes = var_3495_split_sizes_0, x = var_3470_cast_fp16)[name = string("op_3495_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_3497_cast_fp16 = mul(x = var_3495_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3497_cast_fp16")];
            int32 var_3499 = const()[name = string("op_3499"), val = int32(-2)];
            bool var_3500_interleave_0 = const()[name = string("op_3500_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3500_cast_fp16 = concat(axis = var_3499, interleave = var_3500_interleave_0, values = (var_3497_cast_fp16, var_3495_cast_fp16_0))[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3501_cast_fp16 = mul(x = var_3500_cast_fp16, y = var_460_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_95_cast_fp16 = add(x = var_3494_cast_fp16, y = var_3501_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_16)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_18_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_18 = read_state(input = key_cache)[name = string("coreml_update_state_18")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3477_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_17)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_19_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_19 = read_state(input = value_cache)[name = string("coreml_update_state_19")];
            tensor<int32, [4]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = coreml_update_state_18)[name = string("op_3571_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3574_axis_0 = const()[name = string("op_3574_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3574_cast_fp16_1 = split(axis = var_3574_axis_0, split_sizes = tile_18, x = var_3571_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3581_begin_0 = const()[name = string("op_3581_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3581_end_0 = const()[name = string("op_3581_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3581_end_mask_0 = const()[name = string("op_3581_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3581_cast_fp16 = slice_by_index(begin = var_3581_begin_0, end = var_3581_end_0, end_mask = var_3581_end_mask_0, x = coreml_update_state_19)[name = string("op_3581_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3584_axis_0 = const()[name = string("op_3584_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3584_cast_fp16_1 = split(axis = var_3584_axis_0, split_sizes = tile_19, x = var_3581_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [2]> var_3587_split_sizes_0 = const()[name = string("op_3587_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3587_axis_0 = const()[name = string("op_3587_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3587_0, tensor<fp16, [1, 8, 128, 1]> var_3587_1 = split(axis = var_3587_axis_0, split_sizes = var_3587_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3587")];
            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_3574_cast_fp16_0, y = var_3587_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_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_3590_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_3594 = const()[name = string("op_3594"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_151_cast_fp16 = softmax(axis = var_3594, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3600_transpose_x_1 = const()[name = string("op_3600_transpose_x_1"), val = bool(true)];
            bool var_3600_transpose_y_1 = const()[name = string("op_3600_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3600_cast_fp16 = matmul(transpose_x = var_3600_transpose_x_1, transpose_y = var_3600_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3584_cast_fp16_0)[name = string("op_3600_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_3574_cast_fp16_1, y = var_3587_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_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_3602_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_3606 = const()[name = string("op_3606"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_159_cast_fp16 = softmax(axis = var_3606, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3584_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3614 = const()[name = string("op_3614"), 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_3614, interleave = attn_output_75_interleave_0, values = (var_3600_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3618_perm_0 = const()[name = string("op_3618_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_3618_cast_fp16 = transpose(perm = var_3618_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3618_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3651_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3651_cast_fp16")];
            int32 var_3649 = const()[name = string("op_3649"), 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_3649, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3651_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640969472)))];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3661_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3672_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3672_cast_fp16_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3689_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [2]> var_3695_strides_0 = const()[name = string("op_3695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3695_pad_type_0 = const()[name = string("op_3695_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3695_pad_0 = const()[name = string("op_3695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3695_dilations_0 = const()[name = string("op_3695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3695_groups_0 = const()[name = string("op_3695_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3695_cast_fp16 = conv(dilations = var_3695_dilations_0, groups = var_3695_groups_0, pad = var_3695_pad_0, pad_type = var_3695_pad_type_0, strides = var_3695_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3672_cast_fp16_0)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_99_cast_fp16 = mul(x = var_3689_cast_fp16, y = var_3695_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3713_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3713_cast_fp16")];
            int32 var_3711 = const()[name = string("op_3711"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_81_cast_fp16 = concat(axis = var_3711, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3713_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640977728)))];
            fp16 var_3723_to_fp16 = const()[name = string("op_3723_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3723_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3734_split_sizes_0 = const()[name = string("op_3734_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3734_axis_0 = const()[name = string("op_3734_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3734_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3734_cast_fp16_1 = split(axis = var_3734_axis_0, split_sizes = var_3734_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_cast_fp16, x = var_3734_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640985984)))];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = var_3734_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3791_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3798_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3802_cast_fp16 = mul(x = x_101_cast_fp16, y = var_453_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [2]> var_3803_split_sizes_0 = const()[name = string("op_3803_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3803_axis_0 = const()[name = string("op_3803_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3803_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3803_cast_fp16_1 = split(axis = var_3803_axis_0, split_sizes = var_3803_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3805_cast_fp16 = mul(x = var_3803_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3805_cast_fp16")];
            int32 var_3807 = const()[name = string("op_3807"), val = int32(-2)];
            bool var_3808_interleave_0 = const()[name = string("op_3808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3808_cast_fp16 = concat(axis = var_3807, interleave = var_3808_interleave_0, values = (var_3805_cast_fp16, var_3803_cast_fp16_0))[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3809_cast_fp16 = mul(x = var_3808_cast_fp16, y = var_460_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_63_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3809_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3815_cast_fp16 = mul(x = var_3791_cast_fp16, y = var_453_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [2]> var_3816_split_sizes_0 = const()[name = string("op_3816_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3816_axis_0 = const()[name = string("op_3816_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3816_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3816_cast_fp16_1 = split(axis = var_3816_axis_0, split_sizes = var_3816_split_sizes_0, x = var_3791_cast_fp16)[name = string("op_3816_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3818_cast_fp16 = mul(x = var_3816_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3818_cast_fp16")];
            int32 var_3820 = const()[name = string("op_3820"), val = int32(-2)];
            bool var_3821_interleave_0 = const()[name = string("op_3821_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3821_cast_fp16 = concat(axis = var_3820, interleave = var_3821_interleave_0, values = (var_3818_cast_fp16, var_3816_cast_fp16_0))[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3822_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_460_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_105_cast_fp16 = add(x = var_3815_cast_fp16, y = var_3822_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_18)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_20_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_20 = read_state(input = key_cache)[name = string("coreml_update_state_20")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3798_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_19)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_21_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_21 = read_state(input = value_cache)[name = string("coreml_update_state_21")];
            tensor<int32, [4]> var_3892_begin_0 = const()[name = string("op_3892_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3892_end_0 = const()[name = string("op_3892_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3892_end_mask_0 = const()[name = string("op_3892_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3892_cast_fp16 = slice_by_index(begin = var_3892_begin_0, end = var_3892_end_0, end_mask = var_3892_end_mask_0, x = coreml_update_state_20)[name = string("op_3892_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3895_axis_0 = const()[name = string("op_3895_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3895_cast_fp16_1 = split(axis = var_3895_axis_0, split_sizes = tile_20, x = var_3892_cast_fp16)[name = string("op_3895_cast_fp16")];
            tensor<int32, [4]> var_3902_begin_0 = const()[name = string("op_3902_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3902_end_0 = const()[name = string("op_3902_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3902_end_mask_0 = const()[name = string("op_3902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3902_cast_fp16 = slice_by_index(begin = var_3902_begin_0, end = var_3902_end_0, end_mask = var_3902_end_mask_0, x = coreml_update_state_21)[name = string("op_3902_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = tile_21, x = var_3902_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<int32, [2]> var_3908_split_sizes_0 = const()[name = string("op_3908_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3908_axis_0 = const()[name = string("op_3908_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3908_0, tensor<fp16, [1, 8, 128, 1]> var_3908_1 = split(axis = var_3908_axis_0, split_sizes = var_3908_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3908")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3895_cast_fp16_0, y = var_3908_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3911_to_fp16 = const()[name = string("op_3911_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3911_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3915 = const()[name = string("op_3915"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_167_cast_fp16 = softmax(axis = var_3915, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3921_transpose_x_1 = const()[name = string("op_3921_transpose_x_1"), val = bool(true)];
            bool var_3921_transpose_y_1 = const()[name = string("op_3921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3921_cast_fp16 = matmul(transpose_x = var_3921_transpose_x_1, transpose_y = var_3921_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3905_cast_fp16_0)[name = string("op_3921_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3895_cast_fp16_1, y = var_3908_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3923_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3927 = const()[name = string("op_3927"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_175_cast_fp16 = softmax(axis = var_3927, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3905_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_83_cast_fp16 = concat(axis = var_3935, interleave = attn_output_83_interleave_0, values = (var_3921_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3939_perm_0 = const()[name = string("op_3939_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3939_cast_fp16 = transpose(perm = var_3939_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3939_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3972_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3972_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_85_cast_fp16 = concat(axis = var_3970, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3972_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642034624)))];
            fp16 var_3982_to_fp16 = const()[name = string("op_3982_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3982_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3993_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3993_cast_fp16_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3993_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4010_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<int32, [2]> var_4016_strides_0 = const()[name = string("op_4016_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4016_pad_type_0 = const()[name = string("op_4016_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4016_pad_0 = const()[name = string("op_4016_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4016_dilations_0 = const()[name = string("op_4016_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4016_groups_0 = const()[name = string("op_4016_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4016_cast_fp16 = conv(dilations = var_4016_dilations_0, groups = var_4016_groups_0, pad = var_4016_pad_0, pad_type = var_4016_pad_type_0, strides = var_4016_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3993_cast_fp16_0)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_109_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4016_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4034_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4032 = const()[name = string("op_4032"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_89_cast_fp16 = concat(axis = var_4032, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4034_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642042880)))];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4044_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4055_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4055_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4055_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642051136)))];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = var_4055_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4112_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4119_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4119_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4123_cast_fp16 = mul(x = x_111_cast_fp16, y = var_453_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<int32, [2]> var_4124_split_sizes_0 = const()[name = string("op_4124_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4124_axis_0 = const()[name = string("op_4124_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4124_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4124_cast_fp16_1 = split(axis = var_4124_axis_0, split_sizes = var_4124_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4124_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4126_cast_fp16 = mul(x = var_4124_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4126_cast_fp16")];
            int32 var_4128 = const()[name = string("op_4128"), val = int32(-2)];
            bool var_4129_interleave_0 = const()[name = string("op_4129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4129_cast_fp16 = concat(axis = var_4128, interleave = var_4129_interleave_0, values = (var_4126_cast_fp16, var_4124_cast_fp16_0))[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_460_cast_fp16)[name = string("op_4130_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_69_cast_fp16 = add(x = var_4123_cast_fp16, y = var_4130_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4136_cast_fp16 = mul(x = var_4112_cast_fp16, y = var_453_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [2]> var_4137_split_sizes_0 = const()[name = string("op_4137_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4137_axis_0 = const()[name = string("op_4137_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4137_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4137_cast_fp16_1 = split(axis = var_4137_axis_0, split_sizes = var_4137_split_sizes_0, x = var_4112_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4139_cast_fp16 = mul(x = var_4137_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4139_cast_fp16")];
            int32 var_4141 = const()[name = string("op_4141"), val = int32(-2)];
            bool var_4142_interleave_0 = const()[name = string("op_4142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4142_cast_fp16 = concat(axis = var_4141, interleave = var_4142_interleave_0, values = (var_4139_cast_fp16, var_4137_cast_fp16_0))[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = var_460_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_115_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4143_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_20)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_22_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_22 = read_state(input = key_cache)[name = string("coreml_update_state_22")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4119_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_21)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_23_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_23 = read_state(input = value_cache)[name = string("coreml_update_state_23")];
            tensor<int32, [4]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4213_end_0 = const()[name = string("op_4213_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = var_4213_end_0, end_mask = var_4213_end_mask_0, x = coreml_update_state_22)[name = string("op_4213_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4216_axis_0 = const()[name = string("op_4216_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4216_cast_fp16_1 = split(axis = var_4216_axis_0, split_sizes = tile_22, x = var_4213_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, x = coreml_update_state_23)[name = string("op_4223_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4226_axis_0 = const()[name = string("op_4226_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4226_cast_fp16_1 = split(axis = var_4226_axis_0, split_sizes = tile_23, x = var_4223_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [2]> var_4229_split_sizes_0 = const()[name = string("op_4229_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4229_axis_0 = const()[name = string("op_4229_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4229_0, tensor<fp16, [1, 8, 128, 1]> var_4229_1 = split(axis = var_4229_axis_0, split_sizes = var_4229_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4229")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4216_cast_fp16_0, y = var_4229_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4232_to_fp16 = const()[name = string("op_4232_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4232_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4236 = const()[name = string("op_4236"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_183_cast_fp16 = softmax(axis = var_4236, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4242_transpose_x_1 = const()[name = string("op_4242_transpose_x_1"), val = bool(true)];
            bool var_4242_transpose_y_1 = const()[name = string("op_4242_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4242_cast_fp16 = matmul(transpose_x = var_4242_transpose_x_1, transpose_y = var_4242_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4226_cast_fp16_0)[name = string("op_4242_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4216_cast_fp16_1, y = var_4229_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4244_to_fp16 = const()[name = string("op_4244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4244_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4248 = const()[name = string("op_4248"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_191_cast_fp16 = softmax(axis = var_4248, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4226_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4256 = const()[name = string("op_4256"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_91_cast_fp16 = concat(axis = var_4256, interleave = attn_output_91_interleave_0, values = (var_4242_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4260_perm_0 = const()[name = string("op_4260_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4260_cast_fp16 = transpose(perm = var_4260_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4260_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4293_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4293_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_93_cast_fp16 = concat(axis = var_4291, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4293_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643099776)))];
            fp16 var_4303_to_fp16 = const()[name = string("op_4303_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4303_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4314_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4314_cast_fp16_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4331_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4331_cast_fp16")];
            tensor<int32, [2]> var_4337_strides_0 = const()[name = string("op_4337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4337_pad_type_0 = const()[name = string("op_4337_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4337_pad_0 = const()[name = string("op_4337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4337_dilations_0 = const()[name = string("op_4337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4337_groups_0 = const()[name = string("op_4337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4337_cast_fp16 = conv(dilations = var_4337_dilations_0, groups = var_4337_groups_0, pad = var_4337_pad_0, pad_type = var_4337_pad_type_0, strides = var_4337_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4314_cast_fp16_0)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_119_cast_fp16 = mul(x = var_4331_cast_fp16, y = var_4337_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643108032)))];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4355_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4355_cast_fp16")];
            int32 var_4353 = const()[name = string("op_4353"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_97_cast_fp16 = concat(axis = var_4353, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4355_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668273920)))];
            fp16 var_4365_to_fp16 = const()[name = string("op_4365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4365_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4376_split_sizes_0 = const()[name = string("op_4376_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4376_axis_0 = const()[name = string("op_4376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4376_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4376_cast_fp16_1 = split(axis = var_4376_axis_0, split_sizes = var_4376_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4376_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668282176)))];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = var_4376_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4433_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4440_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4444_cast_fp16 = mul(x = x_121_cast_fp16, y = var_453_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<int32, [2]> var_4445_split_sizes_0 = const()[name = string("op_4445_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4445_axis_0 = const()[name = string("op_4445_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4445_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4445_cast_fp16_1 = split(axis = var_4445_axis_0, split_sizes = var_4445_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4445_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4447_cast_fp16 = mul(x = var_4445_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4447_cast_fp16")];
            int32 var_4449 = const()[name = string("op_4449"), val = int32(-2)];
            bool var_4450_interleave_0 = const()[name = string("op_4450_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4450_cast_fp16 = concat(axis = var_4449, interleave = var_4450_interleave_0, values = (var_4447_cast_fp16, var_4445_cast_fp16_0))[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4451_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_460_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_75_cast_fp16 = add(x = var_4444_cast_fp16, y = var_4451_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4457_cast_fp16 = mul(x = var_4433_cast_fp16, y = var_453_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [2]> var_4458_split_sizes_0 = const()[name = string("op_4458_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4458_axis_0 = const()[name = string("op_4458_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4458_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4458_cast_fp16_1 = split(axis = var_4458_axis_0, split_sizes = var_4458_split_sizes_0, x = var_4433_cast_fp16)[name = string("op_4458_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4460_cast_fp16 = mul(x = var_4458_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4460_cast_fp16")];
            int32 var_4462 = const()[name = string("op_4462"), val = int32(-2)];
            bool var_4463_interleave_0 = const()[name = string("op_4463_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4463_cast_fp16 = concat(axis = var_4462, interleave = var_4463_interleave_0, values = (var_4460_cast_fp16, var_4458_cast_fp16_0))[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4464_cast_fp16 = mul(x = var_4463_cast_fp16, y = var_460_cast_fp16)[name = string("op_4464_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_125_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4464_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_22)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_24_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_24 = read_state(input = key_cache)[name = string("coreml_update_state_24")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4440_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_23)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_25_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_25 = read_state(input = value_cache)[name = string("coreml_update_state_25")];
            tensor<int32, [4]> var_4534_begin_0 = const()[name = string("op_4534_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4534_end_0 = const()[name = string("op_4534_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4534_end_mask_0 = const()[name = string("op_4534_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4534_cast_fp16 = slice_by_index(begin = var_4534_begin_0, end = var_4534_end_0, end_mask = var_4534_end_mask_0, x = coreml_update_state_24)[name = string("op_4534_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4537_axis_0 = const()[name = string("op_4537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4537_cast_fp16_1 = split(axis = var_4537_axis_0, split_sizes = tile_24, x = var_4534_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4544_begin_0 = const()[name = string("op_4544_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4544_end_0 = const()[name = string("op_4544_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4544_end_mask_0 = const()[name = string("op_4544_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4544_cast_fp16 = slice_by_index(begin = var_4544_begin_0, end = var_4544_end_0, end_mask = var_4544_end_mask_0, x = coreml_update_state_25)[name = string("op_4544_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4547_axis_0 = const()[name = string("op_4547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4547_cast_fp16_1 = split(axis = var_4547_axis_0, split_sizes = tile_25, x = var_4544_cast_fp16)[name = string("op_4547_cast_fp16")];
            tensor<int32, [2]> var_4550_split_sizes_0 = const()[name = string("op_4550_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4550_axis_0 = const()[name = string("op_4550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4550_0, tensor<fp16, [1, 8, 128, 1]> var_4550_1 = split(axis = var_4550_axis_0, split_sizes = var_4550_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4550")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4537_cast_fp16_0, y = var_4550_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4553_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4557 = const()[name = string("op_4557"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_199_cast_fp16 = softmax(axis = var_4557, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4563_transpose_x_1 = const()[name = string("op_4563_transpose_x_1"), val = bool(true)];
            bool var_4563_transpose_y_1 = const()[name = string("op_4563_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4563_cast_fp16 = matmul(transpose_x = var_4563_transpose_x_1, transpose_y = var_4563_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4547_cast_fp16_0)[name = string("op_4563_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4537_cast_fp16_1, y = var_4550_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4569 = const()[name = string("op_4569"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_207_cast_fp16 = softmax(axis = var_4569, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4547_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4577 = const()[name = string("op_4577"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_99_cast_fp16 = concat(axis = var_4577, interleave = attn_output_99_interleave_0, values = (var_4563_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4581_perm_0 = const()[name = string("op_4581_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4581_cast_fp16 = transpose(perm = var_4581_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4581_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4614_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4614_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_101_cast_fp16 = concat(axis = var_4612, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4614_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669330816)))];
            fp16 var_4624_to_fp16 = const()[name = string("op_4624_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4624_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4635_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4635_cast_fp16_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4635_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4652_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669339072)))];
            tensor<int32, [2]> var_4658_strides_0 = const()[name = string("op_4658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4658_pad_type_0 = const()[name = string("op_4658_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4658_pad_0 = const()[name = string("op_4658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4658_dilations_0 = const()[name = string("op_4658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4658_groups_0 = const()[name = string("op_4658_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4658_cast_fp16 = conv(dilations = var_4658_dilations_0, groups = var_4658_groups_0, pad = var_4658_pad_0, pad_type = var_4658_pad_type_0, strides = var_4658_strides_0, weight = layers_12_mlp_up_proj_weight_to_fp16, x = var_4635_cast_fp16_0)[name = string("op_4658_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_129_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4658_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694504960)))];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4676_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4676_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_105_cast_fp16 = concat(axis = var_4674, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4676_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719670848)))];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4686_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4697_split_sizes_0 = const()[name = string("op_4697_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4697_axis_0 = const()[name = string("op_4697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4697_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4697_cast_fp16_1 = split(axis = var_4697_axis_0, split_sizes = var_4697_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719679104)))];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728067776)))];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729116416)))];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = var_4697_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4754_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4761_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_453_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [2]> var_4766_split_sizes_0 = const()[name = string("op_4766_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4766_axis_0 = const()[name = string("op_4766_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4766_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4766_cast_fp16_1 = split(axis = var_4766_axis_0, split_sizes = var_4766_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4766_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4768_cast_fp16 = mul(x = var_4766_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4768_cast_fp16")];
            int32 var_4770 = const()[name = string("op_4770"), val = int32(-2)];
            bool var_4771_interleave_0 = const()[name = string("op_4771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4771_cast_fp16 = concat(axis = var_4770, interleave = var_4771_interleave_0, values = (var_4768_cast_fp16, var_4766_cast_fp16_0))[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = var_460_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_81_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4778_cast_fp16 = mul(x = var_4754_cast_fp16, y = var_453_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<int32, [2]> var_4779_split_sizes_0 = const()[name = string("op_4779_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4779_axis_0 = const()[name = string("op_4779_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4779_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4779_cast_fp16_1 = split(axis = var_4779_axis_0, split_sizes = var_4779_split_sizes_0, x = var_4754_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4781_cast_fp16")];
            int32 var_4783 = const()[name = string("op_4783"), val = int32(-2)];
            bool var_4784_interleave_0 = const()[name = string("op_4784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4784_cast_fp16 = concat(axis = var_4783, interleave = var_4784_interleave_0, values = (var_4781_cast_fp16, var_4779_cast_fp16_0))[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4785_cast_fp16 = mul(x = var_4784_cast_fp16, y = var_460_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_135_cast_fp16 = add(x = var_4778_cast_fp16, y = var_4785_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_24)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_26_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_26 = read_state(input = key_cache)[name = string("coreml_update_state_26")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4761_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_25)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_27_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_27 = read_state(input = value_cache)[name = string("coreml_update_state_27")];
            tensor<int32, [4]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = coreml_update_state_26)[name = string("op_4855_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4858_axis_0 = const()[name = string("op_4858_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4858_cast_fp16_1 = split(axis = var_4858_axis_0, split_sizes = tile_26, x = var_4855_cast_fp16)[name = string("op_4858_cast_fp16")];
            tensor<int32, [4]> var_4865_begin_0 = const()[name = string("op_4865_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4865_end_0 = const()[name = string("op_4865_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4865_end_mask_0 = const()[name = string("op_4865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4865_cast_fp16 = slice_by_index(begin = var_4865_begin_0, end = var_4865_end_0, end_mask = var_4865_end_mask_0, x = coreml_update_state_27)[name = string("op_4865_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4868_axis_0 = const()[name = string("op_4868_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4868_cast_fp16_1 = split(axis = var_4868_axis_0, split_sizes = tile_27, x = var_4865_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<int32, [2]> var_4871_split_sizes_0 = const()[name = string("op_4871_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4871_axis_0 = const()[name = string("op_4871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4871_0, tensor<fp16, [1, 8, 128, 1]> var_4871_1 = split(axis = var_4871_axis_0, split_sizes = var_4871_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4871")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4858_cast_fp16_0, y = var_4871_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4874_to_fp16 = const()[name = string("op_4874_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4874_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4878 = const()[name = string("op_4878"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_215_cast_fp16 = softmax(axis = var_4878, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4884_transpose_x_1 = const()[name = string("op_4884_transpose_x_1"), val = bool(true)];
            bool var_4884_transpose_y_1 = const()[name = string("op_4884_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4884_cast_fp16 = matmul(transpose_x = var_4884_transpose_x_1, transpose_y = var_4884_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4868_cast_fp16_0)[name = string("op_4884_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4858_cast_fp16_1, y = var_4871_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4886_to_fp16 = const()[name = string("op_4886_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4886_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4890 = const()[name = string("op_4890"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_4890, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4868_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_107_cast_fp16 = concat(axis = var_4898, interleave = attn_output_107_interleave_0, values = (var_4884_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4902_perm_0 = const()[name = string("op_4902_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4902_cast_fp16 = transpose(perm = var_4902_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4902_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730165056)))];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4935_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4935_cast_fp16")];
            int32 var_4933 = const()[name = string("op_4933"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_109_cast_fp16 = concat(axis = var_4933, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4935_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738553728)))];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_4945_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4956_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4956_cast_fp16_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = out_55_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738561984)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4973_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4973_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763727872)))];
            tensor<int32, [2]> var_4979_strides_0 = const()[name = string("op_4979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4979_pad_type_0 = const()[name = string("op_4979_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4979_pad_0 = const()[name = string("op_4979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4979_dilations_0 = const()[name = string("op_4979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4979_groups_0 = const()[name = string("op_4979_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4979_cast_fp16 = conv(dilations = var_4979_dilations_0, groups = var_4979_groups_0, pad = var_4979_pad_0, pad_type = var_4979_pad_type_0, strides = var_4979_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4956_cast_fp16_0)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_4973_cast_fp16, y = var_4979_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788893760)))];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4997_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4997_cast_fp16")];
            int32 var_4995 = const()[name = string("op_4995"), val = int32(1)];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_113_cast_fp16 = concat(axis = var_4995, interleave = doubled_113_interleave_0, values = (hidden_states_cast_fp16, var_4997_cast_fp16))[name = string("doubled_113_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(814059648)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5007_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_5018_split_sizes_0 = const()[name = string("op_5018_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5018_axis_0 = const()[name = string("op_5018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states, tensor<fp16, [1, 2048, 1, 1]> var_5018_cast_fp16_1 = split(axis = var_5018_axis_0, split_sizes = var_5018_split_sizes_0, x = out_cast_fp16)[name = string("op_5018_cast_fp16")];
        } -> (hidden_states);
}