diff --git "a/CodeDecoder.mlmodelc/model.mil" "b/CodeDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/CodeDecoder.mlmodelc/model.mil" @@ -0,0 +1,5444 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache_length, tensor input_embeds, tensor key_cache, tensor key_padding_mask, tensor kv_cache_update_mask, tensor value_cache) { + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("fp32")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("fp32")]; + string cast_2_dtype_0 = const()[name = string("cast_2_dtype_0"), val = string("fp32")]; + string cast_3_dtype_0 = const()[name = string("cast_3_dtype_0"), val = string("fp32")]; + string cast_4_dtype_0 = const()[name = string("cast_4_dtype_0"), val = string("fp32")]; + tensor layers_0_self_attn_q_proj_weight = const()[name = string("layers_0_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor layers_0_self_attn_k_proj_weight = const()[name = string("layers_0_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8388736)))]; + tensor layers_0_self_attn_v_proj_weight = const()[name = string("layers_0_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12583104)))]; + tensor layers_0_self_attn_o_proj_weight = const()[name = string("layers_0_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16777472)))]; + tensor layers_0_mlp_gate_proj_weight = const()[name = string("layers_0_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166144)))]; + tensor layers_0_mlp_up_proj_weight = const()[name = string("layers_0_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37749120)))]; + tensor layers_0_mlp_down_proj_weight = const()[name = string("layers_0_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50332096)))]; + tensor layers_1_self_attn_q_proj_weight = const()[name = string("layers_1_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62915072)))]; + tensor layers_1_self_attn_k_proj_weight = const()[name = string("layers_1_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71303744)))]; + tensor layers_1_self_attn_v_proj_weight = const()[name = string("layers_1_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75498112)))]; + tensor layers_1_self_attn_o_proj_weight = const()[name = string("layers_1_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79692480)))]; + tensor layers_1_mlp_gate_proj_weight = const()[name = string("layers_1_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88081152)))]; + tensor layers_1_mlp_up_proj_weight = const()[name = string("layers_1_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100664128)))]; + tensor layers_1_mlp_down_proj_weight = const()[name = string("layers_1_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113247104)))]; + tensor layers_2_self_attn_q_proj_weight = const()[name = string("layers_2_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125830080)))]; + tensor layers_2_self_attn_k_proj_weight = const()[name = string("layers_2_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134218752)))]; + tensor layers_2_self_attn_v_proj_weight = const()[name = string("layers_2_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138413120)))]; + tensor layers_2_self_attn_o_proj_weight = const()[name = string("layers_2_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142607488)))]; + tensor layers_2_mlp_gate_proj_weight = const()[name = string("layers_2_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150996160)))]; + tensor layers_2_mlp_up_proj_weight = const()[name = string("layers_2_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163579136)))]; + tensor layers_2_mlp_down_proj_weight = const()[name = string("layers_2_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176162112)))]; + tensor layers_3_self_attn_q_proj_weight = const()[name = string("layers_3_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188745088)))]; + tensor layers_3_self_attn_k_proj_weight = const()[name = string("layers_3_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197133760)))]; + tensor layers_3_self_attn_v_proj_weight = const()[name = string("layers_3_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201328128)))]; + tensor layers_3_self_attn_o_proj_weight = const()[name = string("layers_3_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205522496)))]; + tensor layers_3_mlp_gate_proj_weight = const()[name = string("layers_3_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213911168)))]; + tensor layers_3_mlp_up_proj_weight = const()[name = string("layers_3_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226494144)))]; + tensor layers_3_mlp_down_proj_weight = const()[name = string("layers_3_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239077120)))]; + tensor layers_4_self_attn_q_proj_weight = const()[name = string("layers_4_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251660096)))]; + tensor layers_4_self_attn_k_proj_weight = const()[name = string("layers_4_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260048768)))]; + tensor layers_4_self_attn_v_proj_weight = const()[name = string("layers_4_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264243136)))]; + tensor layers_4_self_attn_o_proj_weight = const()[name = string("layers_4_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268437504)))]; + tensor layers_4_mlp_gate_proj_weight = const()[name = string("layers_4_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276826176)))]; + tensor layers_4_mlp_up_proj_weight = const()[name = string("layers_4_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289409152)))]; + tensor layers_4_mlp_down_proj_weight = const()[name = string("layers_4_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301992128)))]; + tensor layers_5_self_attn_q_proj_weight = const()[name = string("layers_5_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314575104)))]; + tensor layers_5_self_attn_k_proj_weight = const()[name = string("layers_5_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322963776)))]; + tensor layers_5_self_attn_v_proj_weight = const()[name = string("layers_5_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327158144)))]; + tensor layers_5_self_attn_o_proj_weight = const()[name = string("layers_5_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331352512)))]; + tensor layers_5_mlp_gate_proj_weight = const()[name = string("layers_5_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339741184)))]; + tensor layers_5_mlp_up_proj_weight = const()[name = string("layers_5_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352324160)))]; + tensor layers_5_mlp_down_proj_weight = const()[name = string("layers_5_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364907136)))]; + tensor layers_6_self_attn_q_proj_weight = const()[name = string("layers_6_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377490112)))]; + tensor layers_6_self_attn_k_proj_weight = const()[name = string("layers_6_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385878784)))]; + tensor layers_6_self_attn_v_proj_weight = const()[name = string("layers_6_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390073152)))]; + tensor layers_6_self_attn_o_proj_weight = const()[name = string("layers_6_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394267520)))]; + tensor layers_6_mlp_gate_proj_weight = const()[name = string("layers_6_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402656192)))]; + tensor layers_6_mlp_up_proj_weight = const()[name = string("layers_6_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415239168)))]; + tensor layers_6_mlp_down_proj_weight = const()[name = string("layers_6_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427822144)))]; + tensor layers_7_self_attn_q_proj_weight = const()[name = string("layers_7_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440405120)))]; + tensor layers_7_self_attn_k_proj_weight = const()[name = string("layers_7_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448793792)))]; + tensor layers_7_self_attn_v_proj_weight = const()[name = string("layers_7_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452988160)))]; + tensor layers_7_self_attn_o_proj_weight = const()[name = string("layers_7_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457182528)))]; + tensor layers_7_mlp_gate_proj_weight = const()[name = string("layers_7_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465571200)))]; + tensor layers_7_mlp_up_proj_weight = const()[name = string("layers_7_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478154176)))]; + tensor layers_7_mlp_down_proj_weight = const()[name = string("layers_7_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490737152)))]; + tensor layers_8_self_attn_q_proj_weight = const()[name = string("layers_8_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503320128)))]; + tensor layers_8_self_attn_k_proj_weight = const()[name = string("layers_8_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511708800)))]; + tensor layers_8_self_attn_v_proj_weight = const()[name = string("layers_8_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515903168)))]; + tensor layers_8_self_attn_o_proj_weight = const()[name = string("layers_8_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520097536)))]; + tensor layers_8_mlp_gate_proj_weight = const()[name = string("layers_8_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528486208)))]; + tensor layers_8_mlp_up_proj_weight = const()[name = string("layers_8_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541069184)))]; + tensor layers_8_mlp_down_proj_weight = const()[name = string("layers_8_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553652160)))]; + tensor layers_9_self_attn_q_proj_weight = const()[name = string("layers_9_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566235136)))]; + tensor layers_9_self_attn_k_proj_weight = const()[name = string("layers_9_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574623808)))]; + tensor layers_9_self_attn_v_proj_weight = const()[name = string("layers_9_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578818176)))]; + tensor layers_9_self_attn_o_proj_weight = const()[name = string("layers_9_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583012544)))]; + tensor layers_9_mlp_gate_proj_weight = const()[name = string("layers_9_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591401216)))]; + tensor layers_9_mlp_up_proj_weight = const()[name = string("layers_9_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603984192)))]; + tensor layers_9_mlp_down_proj_weight = const()[name = string("layers_9_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616567168)))]; + tensor layers_10_self_attn_q_proj_weight = const()[name = string("layers_10_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629150144)))]; + tensor layers_10_self_attn_k_proj_weight = const()[name = string("layers_10_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637538816)))]; + tensor layers_10_self_attn_v_proj_weight = const()[name = string("layers_10_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(641733184)))]; + tensor layers_10_self_attn_o_proj_weight = const()[name = string("layers_10_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645927552)))]; + tensor layers_10_mlp_gate_proj_weight = const()[name = string("layers_10_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654316224)))]; + tensor layers_10_mlp_up_proj_weight = const()[name = string("layers_10_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666899200)))]; + tensor layers_10_mlp_down_proj_weight = const()[name = string("layers_10_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679482176)))]; + tensor layers_11_self_attn_q_proj_weight = const()[name = string("layers_11_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692065152)))]; + tensor layers_11_self_attn_k_proj_weight = const()[name = string("layers_11_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700453824)))]; + tensor layers_11_self_attn_v_proj_weight = const()[name = string("layers_11_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704648192)))]; + tensor layers_11_self_attn_o_proj_weight = const()[name = string("layers_11_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708842560)))]; + tensor layers_11_mlp_gate_proj_weight = const()[name = string("layers_11_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717231232)))]; + tensor layers_11_mlp_up_proj_weight = const()[name = string("layers_11_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729814208)))]; + tensor layers_11_mlp_down_proj_weight = const()[name = string("layers_11_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(742397184)))]; + tensor layers_12_self_attn_q_proj_weight = const()[name = string("layers_12_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754980160)))]; + tensor layers_12_self_attn_k_proj_weight = const()[name = string("layers_12_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763368832)))]; + tensor layers_12_self_attn_v_proj_weight = const()[name = string("layers_12_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(767563200)))]; + tensor layers_12_self_attn_o_proj_weight = const()[name = string("layers_12_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771757568)))]; + tensor layers_12_mlp_gate_proj_weight = const()[name = string("layers_12_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(780146240)))]; + tensor layers_12_mlp_up_proj_weight = const()[name = string("layers_12_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792729216)))]; + tensor layers_12_mlp_down_proj_weight = const()[name = string("layers_12_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(805312192)))]; + tensor layers_13_self_attn_q_proj_weight = const()[name = string("layers_13_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817895168)))]; + tensor layers_13_self_attn_k_proj_weight = const()[name = string("layers_13_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(826283840)))]; + tensor layers_13_self_attn_v_proj_weight = const()[name = string("layers_13_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(830478208)))]; + tensor layers_13_self_attn_o_proj_weight = const()[name = string("layers_13_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(834672576)))]; + tensor layers_13_mlp_gate_proj_weight = const()[name = string("layers_13_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843061248)))]; + tensor layers_13_mlp_up_proj_weight = const()[name = string("layers_13_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855644224)))]; + tensor layers_13_mlp_down_proj_weight = const()[name = string("layers_13_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(868227200)))]; + tensor layers_14_self_attn_q_proj_weight = const()[name = string("layers_14_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880810176)))]; + tensor layers_14_self_attn_k_proj_weight = const()[name = string("layers_14_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(889198848)))]; + tensor layers_14_self_attn_v_proj_weight = const()[name = string("layers_14_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893393216)))]; + tensor layers_14_self_attn_o_proj_weight = const()[name = string("layers_14_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897587584)))]; + tensor layers_14_mlp_gate_proj_weight = const()[name = string("layers_14_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(905976256)))]; + tensor layers_14_mlp_up_proj_weight = const()[name = string("layers_14_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918559232)))]; + tensor layers_14_mlp_down_proj_weight = const()[name = string("layers_14_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(931142208)))]; + tensor layers_15_self_attn_q_proj_weight = const()[name = string("layers_15_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(943725184)))]; + tensor layers_15_self_attn_k_proj_weight = const()[name = string("layers_15_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(952113856)))]; + tensor layers_15_self_attn_v_proj_weight = const()[name = string("layers_15_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(956308224)))]; + tensor layers_15_self_attn_o_proj_weight = const()[name = string("layers_15_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(960502592)))]; + tensor layers_15_mlp_gate_proj_weight = const()[name = string("layers_15_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(968891264)))]; + tensor layers_15_mlp_up_proj_weight = const()[name = string("layers_15_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(981474240)))]; + tensor layers_15_mlp_down_proj_weight = const()[name = string("layers_15_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(994057216)))]; + tensor layers_16_self_attn_q_proj_weight = const()[name = string("layers_16_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1006640192)))]; + tensor layers_16_self_attn_k_proj_weight = const()[name = string("layers_16_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1015028864)))]; + tensor layers_16_self_attn_v_proj_weight = const()[name = string("layers_16_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1019223232)))]; + tensor layers_16_self_attn_o_proj_weight = const()[name = string("layers_16_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1023417600)))]; + tensor layers_16_mlp_gate_proj_weight = const()[name = string("layers_16_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031806272)))]; + tensor layers_16_mlp_up_proj_weight = const()[name = string("layers_16_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1044389248)))]; + tensor layers_16_mlp_down_proj_weight = const()[name = string("layers_16_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1056972224)))]; + tensor layers_17_self_attn_q_proj_weight = const()[name = string("layers_17_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1069555200)))]; + tensor layers_17_self_attn_k_proj_weight = const()[name = string("layers_17_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1077943872)))]; + tensor layers_17_self_attn_v_proj_weight = const()[name = string("layers_17_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1082138240)))]; + tensor layers_17_self_attn_o_proj_weight = const()[name = string("layers_17_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086332608)))]; + tensor layers_17_mlp_gate_proj_weight = const()[name = string("layers_17_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1094721280)))]; + tensor layers_17_mlp_up_proj_weight = const()[name = string("layers_17_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1107304256)))]; + tensor layers_17_mlp_down_proj_weight = const()[name = string("layers_17_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1119887232)))]; + tensor layers_18_self_attn_q_proj_weight = const()[name = string("layers_18_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1132470208)))]; + tensor layers_18_self_attn_k_proj_weight = const()[name = string("layers_18_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1140858880)))]; + tensor layers_18_self_attn_v_proj_weight = const()[name = string("layers_18_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145053248)))]; + tensor layers_18_self_attn_o_proj_weight = const()[name = string("layers_18_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1149247616)))]; + tensor layers_18_mlp_gate_proj_weight = const()[name = string("layers_18_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1157636288)))]; + tensor layers_18_mlp_up_proj_weight = const()[name = string("layers_18_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1170219264)))]; + tensor layers_18_mlp_down_proj_weight = const()[name = string("layers_18_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1182802240)))]; + tensor layers_19_self_attn_q_proj_weight = const()[name = string("layers_19_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1195385216)))]; + tensor layers_19_self_attn_k_proj_weight = const()[name = string("layers_19_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1203773888)))]; + tensor layers_19_self_attn_v_proj_weight = const()[name = string("layers_19_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1207968256)))]; + tensor layers_19_self_attn_o_proj_weight = const()[name = string("layers_19_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1212162624)))]; + tensor layers_19_mlp_gate_proj_weight = const()[name = string("layers_19_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1220551296)))]; + tensor layers_19_mlp_up_proj_weight = const()[name = string("layers_19_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1233134272)))]; + tensor layers_19_mlp_down_proj_weight = const()[name = string("layers_19_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1245717248)))]; + tensor layers_20_self_attn_q_proj_weight = const()[name = string("layers_20_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1258300224)))]; + tensor layers_20_self_attn_k_proj_weight = const()[name = string("layers_20_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1266688896)))]; + tensor layers_20_self_attn_v_proj_weight = const()[name = string("layers_20_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1270883264)))]; + tensor layers_20_self_attn_o_proj_weight = const()[name = string("layers_20_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1275077632)))]; + tensor layers_20_mlp_gate_proj_weight = const()[name = string("layers_20_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1283466304)))]; + tensor layers_20_mlp_up_proj_weight = const()[name = string("layers_20_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1296049280)))]; + tensor layers_20_mlp_down_proj_weight = const()[name = string("layers_20_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1308632256)))]; + tensor layers_21_self_attn_q_proj_weight = const()[name = string("layers_21_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1321215232)))]; + tensor layers_21_self_attn_k_proj_weight = const()[name = string("layers_21_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1329603904)))]; + tensor layers_21_self_attn_v_proj_weight = const()[name = string("layers_21_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1333798272)))]; + tensor layers_21_self_attn_o_proj_weight = const()[name = string("layers_21_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1337992640)))]; + tensor layers_21_mlp_gate_proj_weight = const()[name = string("layers_21_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1346381312)))]; + tensor layers_21_mlp_up_proj_weight = const()[name = string("layers_21_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1358964288)))]; + tensor layers_21_mlp_down_proj_weight = const()[name = string("layers_21_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1371547264)))]; + tensor layers_22_self_attn_q_proj_weight = const()[name = string("layers_22_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1384130240)))]; + tensor layers_22_self_attn_k_proj_weight = const()[name = string("layers_22_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1392518912)))]; + tensor layers_22_self_attn_v_proj_weight = const()[name = string("layers_22_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1396713280)))]; + tensor layers_22_self_attn_o_proj_weight = const()[name = string("layers_22_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1400907648)))]; + tensor layers_22_mlp_gate_proj_weight = const()[name = string("layers_22_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1409296320)))]; + tensor layers_22_mlp_up_proj_weight = const()[name = string("layers_22_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1421879296)))]; + tensor layers_22_mlp_down_proj_weight = const()[name = string("layers_22_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1434462272)))]; + tensor layers_23_self_attn_q_proj_weight = const()[name = string("layers_23_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1447045248)))]; + tensor layers_23_self_attn_k_proj_weight = const()[name = string("layers_23_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1455433920)))]; + tensor layers_23_self_attn_v_proj_weight = const()[name = string("layers_23_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1459628288)))]; + tensor layers_23_self_attn_o_proj_weight = const()[name = string("layers_23_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1463822656)))]; + tensor layers_23_mlp_gate_proj_weight = const()[name = string("layers_23_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1472211328)))]; + tensor layers_23_mlp_up_proj_weight = const()[name = string("layers_23_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1484794304)))]; + tensor layers_23_mlp_down_proj_weight = const()[name = string("layers_23_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1497377280)))]; + tensor layers_24_self_attn_q_proj_weight = const()[name = string("layers_24_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1509960256)))]; + tensor layers_24_self_attn_k_proj_weight = const()[name = string("layers_24_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1518348928)))]; + tensor layers_24_self_attn_v_proj_weight = const()[name = string("layers_24_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1522543296)))]; + tensor layers_24_self_attn_o_proj_weight = const()[name = string("layers_24_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1526737664)))]; + tensor layers_24_mlp_gate_proj_weight = const()[name = string("layers_24_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1535126336)))]; + tensor layers_24_mlp_up_proj_weight = const()[name = string("layers_24_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1547709312)))]; + tensor layers_24_mlp_down_proj_weight = const()[name = string("layers_24_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1560292288)))]; + tensor layers_25_self_attn_q_proj_weight = const()[name = string("layers_25_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1572875264)))]; + tensor layers_25_self_attn_k_proj_weight = const()[name = string("layers_25_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1581263936)))]; + tensor layers_25_self_attn_v_proj_weight = const()[name = string("layers_25_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1585458304)))]; + tensor layers_25_self_attn_o_proj_weight = const()[name = string("layers_25_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1589652672)))]; + tensor layers_25_mlp_gate_proj_weight = const()[name = string("layers_25_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1598041344)))]; + tensor layers_25_mlp_up_proj_weight = const()[name = string("layers_25_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1610624320)))]; + tensor layers_25_mlp_down_proj_weight = const()[name = string("layers_25_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1623207296)))]; + tensor layers_26_self_attn_q_proj_weight = const()[name = string("layers_26_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1635790272)))]; + tensor layers_26_self_attn_k_proj_weight = const()[name = string("layers_26_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1644178944)))]; + tensor layers_26_self_attn_v_proj_weight = const()[name = string("layers_26_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1648373312)))]; + tensor layers_26_self_attn_o_proj_weight = const()[name = string("layers_26_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1652567680)))]; + tensor layers_26_mlp_gate_proj_weight = const()[name = string("layers_26_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1660956352)))]; + tensor layers_26_mlp_up_proj_weight = const()[name = string("layers_26_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1673539328)))]; + tensor layers_26_mlp_down_proj_weight = const()[name = string("layers_26_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1686122304)))]; + tensor layers_27_self_attn_q_proj_weight = const()[name = string("layers_27_self_attn_q_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1698705280)))]; + tensor layers_27_self_attn_k_proj_weight = const()[name = string("layers_27_self_attn_k_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1707093952)))]; + tensor layers_27_self_attn_v_proj_weight = const()[name = string("layers_27_self_attn_v_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1711288320)))]; + tensor layers_27_self_attn_o_proj_weight = const()[name = string("layers_27_self_attn_o_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1715482688)))]; + tensor layers_27_mlp_gate_proj_weight = const()[name = string("layers_27_mlp_gate_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1723871360)))]; + tensor layers_27_mlp_up_proj_weight = const()[name = string("layers_27_mlp_up_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1736454336)))]; + tensor layers_27_mlp_down_proj_weight = const()[name = string("layers_27_mlp_down_proj_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1749037312)))]; + tensor codec_head_weight = const()[name = string("codec_head_weight"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1761620288)))]; + tensor var_935_axes_0 = const()[name = string("op_935_axes_0"), val = tensor([0])]; + tensor var_935 = expand_dims(axes = var_935_axes_0, x = cache_length)[name = string("op_935")]; + string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("fp32")]; + tensor const_0 = const()[name = string("const_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774203264)))]; + tensor var_956_axes_0 = const()[name = string("op_956_axes_0"), val = tensor([-1])]; + tensor position_ids = cast(dtype = position_ids_dtype_0, x = var_935)[name = string("cast_349")]; + tensor var_956 = expand_dims(axes = var_956_axes_0, x = position_ids)[name = string("op_956")]; + bool var_957_transpose_x_0 = const()[name = string("op_957_transpose_x_0"), val = bool(false)]; + bool var_957_transpose_y_0 = const()[name = string("op_957_transpose_y_0"), val = bool(false)]; + tensor var_957 = matmul(transpose_x = var_957_transpose_x_0, transpose_y = var_957_transpose_y_0, x = const_0, y = var_956)[name = string("op_957")]; + tensor freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor([0, 2, 1])]; + int32 var_962 = const()[name = string("op_962"), val = int32(-1)]; + bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)]; + tensor freqs = transpose(perm = freqs_perm_0, x = var_957)[name = string("transpose_112")]; + tensor emb = concat(axis = var_962, interleave = emb_interleave_0, values = (freqs, freqs))[name = string("emb")]; + tensor var_964 = cos(x = emb)[name = string("op_964")]; + tensor var_972 = sin(x = emb)[name = string("op_972")]; + tensor var_989_begin_0 = const()[name = string("op_989_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_989_end_0 = const()[name = string("op_989_end_0"), val = tensor([1, 1024, 1, 256])]; + tensor var_989_end_mask_0 = const()[name = string("op_989_end_mask_0"), val = tensor([true, false, true, true])]; + tensor cast_3 = cast(dtype = cast_3_dtype_0, x = key_cache)[name = string("cast_351")]; + tensor var_989 = slice_by_index(begin = var_989_begin_0, end = var_989_end_0, end_mask = var_989_end_mask_0, x = cast_3)[name = string("op_989")]; + tensor var_1009_begin_0 = const()[name = string("op_1009_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1009_end_0 = const()[name = string("op_1009_end_0"), val = tensor([1, 1024, 1, 256])]; + tensor var_1009_end_mask_0 = const()[name = string("op_1009_end_mask_0"), val = tensor([true, false, true, true])]; + tensor cast_4 = cast(dtype = cast_4_dtype_0, x = value_cache)[name = string("cast_350")]; + tensor var_1009 = slice_by_index(begin = var_1009_begin_0, end = var_1009_end_0, end_mask = var_1009_end_mask_0, x = cast_4)[name = string("op_1009")]; + tensor var_1021_axes_0 = const()[name = string("op_1021_axes_0"), val = tensor([-1])]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = input_embeds)[name = string("cast_354")]; + tensor var_1021 = squeeze(axes = var_1021_axes_0, x = cast_0)[name = string("op_1021")]; + tensor var_1023_axes_0 = const()[name = string("op_1023_axes_0"), val = tensor([-1])]; + tensor var_1023 = squeeze(axes = var_1023_axes_0, x = var_1021)[name = string("op_1023")]; + tensor hidden_states_1_axes_0 = const()[name = string("hidden_states_1_axes_0"), val = tensor([0])]; + tensor hidden_states_1 = expand_dims(axes = hidden_states_1_axes_0, x = var_1023)[name = string("hidden_states_1")]; + fp32 var_1029_promoted = const()[name = string("op_1029_promoted"), val = fp32(0x1p+1)]; + tensor var_1035 = pow(x = hidden_states_1, y = var_1029_promoted)[name = string("op_1035")]; + tensor variance_1_axes_0 = const()[name = string("variance_1_axes_0"), val = tensor([-1])]; + bool variance_1_keep_dims_0 = const()[name = string("variance_1_keep_dims_0"), val = bool(true)]; + tensor variance_1 = reduce_mean(axes = variance_1_axes_0, keep_dims = variance_1_keep_dims_0, x = var_1035)[name = string("variance_1")]; + fp32 var_1038 = const()[name = string("op_1038"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1039 = add(x = variance_1, y = var_1038)[name = string("op_1039")]; + fp32 var_1040_epsilon_0 = const()[name = string("op_1040_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1040 = rsqrt(epsilon = var_1040_epsilon_0, x = var_1039)[name = string("op_1040")]; + tensor hidden_states_5 = mul(x = hidden_states_1, y = var_1040)[name = string("hidden_states_5")]; + tensor const_1 = const()[name = string("const_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774203584)))]; + tensor input_1 = mul(x = const_1, y = hidden_states_5)[name = string("input_1")]; + tensor linear_0_bias_0 = const()[name = string("linear_0_bias_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774207744)))]; + tensor q_1 = linear(bias = linear_0_bias_0, weight = layers_0_self_attn_q_proj_weight, x = input_1)[name = string("linear_0")]; + tensor linear_1_bias_0 = const()[name = string("linear_1_bias_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774216000)))]; + tensor k_1 = linear(bias = linear_1_bias_0, weight = layers_0_self_attn_k_proj_weight, x = input_1)[name = string("linear_1")]; + tensor v_1 = linear(bias = linear_1_bias_0, weight = layers_0_self_attn_v_proj_weight, x = input_1)[name = string("linear_2")]; + tensor var_1057 = const()[name = string("op_1057"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_7 = reshape(shape = var_1057, x = q_1)[name = string("hidden_states_7")]; + tensor var_1063 = const()[name = string("op_1063"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_13 = reshape(shape = var_1063, x = k_1)[name = string("hidden_states_13")]; + tensor var_1069 = const()[name = string("op_1069"), val = tensor([1, 1, 8, 128])]; + tensor v_3 = reshape(shape = var_1069, x = v_1)[name = string("v_3")]; + fp32 var_1074_promoted = const()[name = string("op_1074_promoted"), val = fp32(0x1p+1)]; + tensor var_1080 = pow(x = hidden_states_7, y = var_1074_promoted)[name = string("op_1080")]; + tensor variance_3_axes_0 = const()[name = string("variance_3_axes_0"), val = tensor([-1])]; + bool variance_3_keep_dims_0 = const()[name = string("variance_3_keep_dims_0"), val = bool(true)]; + tensor variance_3 = reduce_mean(axes = variance_3_axes_0, keep_dims = variance_3_keep_dims_0, x = var_1080)[name = string("variance_3")]; + fp32 var_1083 = const()[name = string("op_1083"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1084 = add(x = variance_3, y = var_1083)[name = string("op_1084")]; + fp32 var_1085_epsilon_0 = const()[name = string("op_1085_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1085 = rsqrt(epsilon = var_1085_epsilon_0, x = var_1084)[name = string("op_1085")]; + tensor hidden_states_11 = mul(x = hidden_states_7, y = var_1085)[name = string("hidden_states_11")]; + tensor const_2 = const()[name = string("const_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774220160)))]; + tensor q_3 = mul(x = const_2, y = hidden_states_11)[name = string("q_3")]; + fp32 var_1092_promoted = const()[name = string("op_1092_promoted"), val = fp32(0x1p+1)]; + tensor var_1098 = pow(x = hidden_states_13, y = var_1092_promoted)[name = string("op_1098")]; + tensor variance_5_axes_0 = const()[name = string("variance_5_axes_0"), val = tensor([-1])]; + bool variance_5_keep_dims_0 = const()[name = string("variance_5_keep_dims_0"), val = bool(true)]; + tensor variance_5 = reduce_mean(axes = variance_5_axes_0, keep_dims = variance_5_keep_dims_0, x = var_1098)[name = string("variance_5")]; + fp32 var_1101 = const()[name = string("op_1101"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1102 = add(x = variance_5, y = var_1101)[name = string("op_1102")]; + fp32 var_1103_epsilon_0 = const()[name = string("op_1103_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1103 = rsqrt(epsilon = var_1103_epsilon_0, x = var_1102)[name = string("op_1103")]; + tensor hidden_states_17 = mul(x = hidden_states_13, y = var_1103)[name = string("hidden_states_17")]; + tensor const_3 = const()[name = string("const_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774220736)))]; + tensor k_3 = mul(x = const_3, y = hidden_states_17)[name = string("k_3")]; + tensor q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor cos_3_axes_0 = const()[name = string("cos_3_axes_0"), val = tensor([1])]; + tensor cos_3 = expand_dims(axes = cos_3_axes_0, x = var_964)[name = string("cos_3")]; + tensor sin_3_axes_0 = const()[name = string("sin_3_axes_0"), val = tensor([1])]; + tensor sin_3 = expand_dims(axes = sin_3_axes_0, x = var_972)[name = string("sin_3")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = q_3)[name = string("transpose_111")]; + tensor var_1120 = mul(x = q_5, y = cos_3)[name = string("op_1120")]; + tensor x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_1 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_5)[name = string("x1_1")]; + tensor x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_1 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_5)[name = string("x2_1")]; + fp32 const_6_promoted = const()[name = string("const_6_promoted"), val = fp32(-0x1p+0)]; + tensor var_1141 = mul(x = x2_1, y = const_6_promoted)[name = string("op_1141")]; + int32 var_1143 = const()[name = string("op_1143"), val = int32(-1)]; + bool var_1144_interleave_0 = const()[name = string("op_1144_interleave_0"), val = bool(false)]; + tensor var_1144 = concat(axis = var_1143, interleave = var_1144_interleave_0, values = (var_1141, x1_1))[name = string("op_1144")]; + tensor var_1145 = mul(x = var_1144, y = sin_3)[name = string("op_1145")]; + tensor q_7 = add(x = var_1120, y = var_1145)[name = string("q_7")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = k_3)[name = string("transpose_110")]; + tensor var_1148 = mul(x = k_5, y = cos_3)[name = string("op_1148")]; + tensor x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_3 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_5)[name = string("x1_3")]; + tensor x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_3 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_5)[name = string("x2_3")]; + fp32 const_9_promoted = const()[name = string("const_9_promoted"), val = fp32(-0x1p+0)]; + tensor var_1169 = mul(x = x2_3, y = const_9_promoted)[name = string("op_1169")]; + int32 var_1171 = const()[name = string("op_1171"), val = int32(-1)]; + bool var_1172_interleave_0 = const()[name = string("op_1172_interleave_0"), val = bool(false)]; + tensor var_1172 = concat(axis = var_1171, interleave = var_1172_interleave_0, values = (var_1169, x1_3))[name = string("op_1172")]; + tensor var_1173 = mul(x = var_1172, y = sin_3)[name = string("op_1173")]; + tensor k_7 = add(x = var_1148, y = var_1173)[name = string("k_7")]; + tensor var_1180 = const()[name = string("op_1180"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_1 = reshape(shape = var_1180, x = k_7)[name = string("nk_flat_1")]; + tensor var_1186 = const()[name = string("op_1186"), val = tensor([1, 1024, 1, 1])]; + tensor v_5 = transpose(perm = v_5_perm_0, x = v_3)[name = string("transpose_109")]; + tensor nv_flat_1 = reshape(shape = var_1186, x = v_5)[name = string("nv_flat_1")]; + tensor var_1189_axes_0 = const()[name = string("op_1189_axes_0"), val = tensor([1])]; + tensor cast_2 = cast(dtype = cast_2_dtype_0, x = kv_cache_update_mask)[name = string("cast_352")]; + tensor var_1189 = expand_dims(axes = var_1189_axes_0, x = cast_2)[name = string("op_1189")]; + tensor update_mask_1_axes_0 = const()[name = string("update_mask_1_axes_0"), val = tensor([2])]; + tensor update_mask_1 = expand_dims(axes = update_mask_1_axes_0, x = var_1189)[name = string("update_mask_1")]; + fp32 var_1192 = const()[name = string("op_1192"), val = fp32(0x1p+0)]; + tensor var_1194 = sub(x = var_1192, y = update_mask_1)[name = string("op_1194")]; + tensor var_1195 = mul(x = var_989, y = var_1194)[name = string("op_1195")]; + tensor var_1196 = mul(x = nk_flat_1, y = update_mask_1)[name = string("op_1196")]; + tensor key_cache_5 = add(x = var_1195, y = var_1196)[name = string("key_cache_5")]; + tensor var_1202 = mul(x = var_1009, y = var_1194)[name = string("op_1202")]; + tensor var_1203 = mul(x = nv_flat_1, y = update_mask_1)[name = string("op_1203")]; + tensor value_cache_5 = add(x = var_1202, y = var_1203)[name = string("value_cache_5")]; + tensor kc_1_axes_0 = const()[name = string("kc_1_axes_0"), val = tensor([2])]; + tensor kc_1 = squeeze(axes = kc_1_axes_0, x = key_cache_5)[name = string("kc_1")]; + tensor var_1212 = const()[name = string("op_1212"), val = tensor([1, 8, 128, 256])]; + tensor kc_3 = reshape(shape = var_1212, x = kc_1)[name = string("kc_3")]; + tensor vc_1_axes_0 = const()[name = string("vc_1_axes_0"), val = tensor([2])]; + tensor vc_1 = squeeze(axes = vc_1_axes_0, x = value_cache_5)[name = string("vc_1")]; + tensor var_1220 = const()[name = string("op_1220"), val = tensor([1, 8, 128, 256])]; + tensor vc_3 = reshape(shape = var_1220, x = vc_1)[name = string("vc_3")]; + tensor var_1223_axes_0 = const()[name = string("op_1223_axes_0"), val = tensor([2])]; + tensor var_1223 = expand_dims(axes = var_1223_axes_0, x = kc_3)[name = string("op_1223")]; + tensor var_1231_reps_0 = const()[name = string("op_1231_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_1231 = tile(reps = var_1231_reps_0, x = var_1223)[name = string("op_1231")]; + tensor var_1236 = const()[name = string("op_1236"), val = tensor([1, 16, 128, 256])]; + tensor kc_5 = reshape(shape = var_1236, x = var_1231)[name = string("kc_5")]; + tensor var_1239_axes_0 = const()[name = string("op_1239_axes_0"), val = tensor([2])]; + tensor var_1239 = expand_dims(axes = var_1239_axes_0, x = vc_3)[name = string("op_1239")]; + tensor var_1247_reps_0 = const()[name = string("op_1247_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_1247 = tile(reps = var_1247_reps_0, x = var_1239)[name = string("op_1247")]; + tensor var_1252 = const()[name = string("op_1252"), val = tensor([1, 16, 128, 256])]; + tensor vc_5 = reshape(shape = var_1252, x = var_1247)[name = string("vc_5")]; + bool var_1254_transpose_x_0 = const()[name = string("op_1254_transpose_x_0"), val = bool(false)]; + bool var_1254_transpose_y_0 = const()[name = string("op_1254_transpose_y_0"), val = bool(false)]; + tensor var_1254 = matmul(transpose_x = var_1254_transpose_x_0, transpose_y = var_1254_transpose_y_0, x = q_7, y = kc_5)[name = string("op_1254")]; + fp32 _inversed_attn_weights_1_y_0 = const()[name = string("_inversed_attn_weights_1_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_1 = mul(x = var_1254, y = _inversed_attn_weights_1_y_0)[name = string("_inversed_attn_weights_1")]; + tensor var_1258_axes_0 = const()[name = string("op_1258_axes_0"), val = tensor([1])]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = key_padding_mask)[name = string("cast_353")]; + tensor var_1258 = expand_dims(axes = var_1258_axes_0, x = cast_1)[name = string("op_1258")]; + tensor mask_1_axes_0 = const()[name = string("mask_1_axes_0"), val = tensor([2])]; + tensor mask_1 = expand_dims(axes = mask_1_axes_0, x = var_1258)[name = string("mask_1")]; + tensor attn_weights_3 = add(x = _inversed_attn_weights_1, y = mask_1)[name = string("attn_weights_3")]; + int32 var_1268 = const()[name = string("op_1268"), val = int32(-1)]; + tensor attn_weights_7 = softmax(axis = var_1268, x = attn_weights_3)[name = string("attn_weights_7")]; + bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(false)]; + bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(true)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_7, y = vc_5)[name = string("attn_output_1")]; + tensor var_1277_perm_0 = const()[name = string("op_1277_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1281 = const()[name = string("op_1281"), val = tensor([1, 1, -1])]; + tensor var_1277 = transpose(perm = var_1277_perm_0, x = attn_output_1)[name = string("transpose_108")]; + tensor input_3 = reshape(shape = var_1281, x = var_1277)[name = string("input_3")]; + tensor attn_output_3 = linear(bias = linear_1_bias_0, weight = layers_0_self_attn_o_proj_weight, x = input_3)[name = string("linear_3")]; + tensor var_1287_axes_0 = const()[name = string("op_1287_axes_0"), val = tensor([0])]; + tensor var_1287 = squeeze(axes = var_1287_axes_0, x = attn_output_3)[name = string("op_1287")]; + tensor var_1289_axes_0 = const()[name = string("op_1289_axes_0"), val = tensor([0])]; + tensor var_1289 = squeeze(axes = var_1289_axes_0, x = var_1287)[name = string("op_1289")]; + tensor var_1291_axes_0 = const()[name = string("op_1291_axes_0"), val = tensor([-1])]; + tensor var_1291 = expand_dims(axes = var_1291_axes_0, x = var_1289)[name = string("op_1291")]; + tensor attn_4d_1_axes_0 = const()[name = string("attn_4d_1_axes_0"), val = tensor([-1])]; + tensor attn_4d_1 = expand_dims(axes = attn_4d_1_axes_0, x = var_1291)[name = string("attn_4d_1")]; + tensor hidden_1 = add(x = cast_0, y = attn_4d_1)[name = string("hidden_1")]; + tensor var_1297_axes_0 = const()[name = string("op_1297_axes_0"), val = tensor([-1])]; + tensor var_1297 = squeeze(axes = var_1297_axes_0, x = hidden_1)[name = string("op_1297")]; + tensor var_1299_axes_0 = const()[name = string("op_1299_axes_0"), val = tensor([-1])]; + tensor var_1299 = squeeze(axes = var_1299_axes_0, x = var_1297)[name = string("op_1299")]; + tensor hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor([0])]; + tensor hidden_states_19 = expand_dims(axes = hidden_states_19_axes_0, x = var_1299)[name = string("hidden_states_19")]; + fp32 var_1305_promoted = const()[name = string("op_1305_promoted"), val = fp32(0x1p+1)]; + tensor var_1311 = pow(x = hidden_states_19, y = var_1305_promoted)[name = string("op_1311")]; + tensor variance_7_axes_0 = const()[name = string("variance_7_axes_0"), val = tensor([-1])]; + bool variance_7_keep_dims_0 = const()[name = string("variance_7_keep_dims_0"), val = bool(true)]; + tensor variance_7 = reduce_mean(axes = variance_7_axes_0, keep_dims = variance_7_keep_dims_0, x = var_1311)[name = string("variance_7")]; + fp32 var_1314 = const()[name = string("op_1314"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1315 = add(x = variance_7, y = var_1314)[name = string("op_1315")]; + fp32 var_1316_epsilon_0 = const()[name = string("op_1316_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1316 = rsqrt(epsilon = var_1316_epsilon_0, x = var_1315)[name = string("op_1316")]; + tensor hidden_states_23 = mul(x = hidden_states_19, y = var_1316)[name = string("hidden_states_23")]; + tensor const_10 = const()[name = string("const_10"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774221312)))]; + tensor input_5 = mul(x = const_10, y = hidden_states_23)[name = string("input_5")]; + tensor linear_4_bias_0 = const()[name = string("linear_4_bias_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774225472)))]; + tensor input_7 = linear(bias = linear_4_bias_0, weight = layers_0_mlp_gate_proj_weight, x = input_5)[name = string("linear_4")]; + tensor var_1326 = silu(x = input_7)[name = string("op_1326")]; + tensor var_1328 = linear(bias = linear_4_bias_0, weight = layers_0_mlp_up_proj_weight, x = input_5)[name = string("linear_5")]; + tensor input_9 = mul(x = var_1326, y = var_1328)[name = string("input_9")]; + tensor mlp_out_1 = linear(bias = linear_1_bias_0, weight = layers_0_mlp_down_proj_weight, x = input_9)[name = string("linear_6")]; + tensor var_1333_axes_0 = const()[name = string("op_1333_axes_0"), val = tensor([0])]; + tensor var_1333 = squeeze(axes = var_1333_axes_0, x = mlp_out_1)[name = string("op_1333")]; + tensor var_1335_axes_0 = const()[name = string("op_1335_axes_0"), val = tensor([0])]; + tensor var_1335 = squeeze(axes = var_1335_axes_0, x = var_1333)[name = string("op_1335")]; + tensor var_1337_axes_0 = const()[name = string("op_1337_axes_0"), val = tensor([-1])]; + tensor var_1337 = expand_dims(axes = var_1337_axes_0, x = var_1335)[name = string("op_1337")]; + tensor mlp_4d_1_axes_0 = const()[name = string("mlp_4d_1_axes_0"), val = tensor([-1])]; + tensor mlp_4d_1 = expand_dims(axes = mlp_4d_1_axes_0, x = var_1337)[name = string("mlp_4d_1")]; + tensor hidden_3 = add(x = hidden_1, y = mlp_4d_1)[name = string("hidden_3")]; + tensor var_1351_begin_0 = const()[name = string("op_1351_begin_0"), val = tensor([0, 1024, 0, 0])]; + tensor var_1351_end_0 = const()[name = string("op_1351_end_0"), val = tensor([1, 2048, 1, 256])]; + tensor var_1351_end_mask_0 = const()[name = string("op_1351_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1351 = slice_by_index(begin = var_1351_begin_0, end = var_1351_end_0, end_mask = var_1351_end_mask_0, x = cast_3)[name = string("op_1351")]; + tensor var_1371_begin_0 = const()[name = string("op_1371_begin_0"), val = tensor([0, 1024, 0, 0])]; + tensor var_1371_end_0 = const()[name = string("op_1371_end_0"), val = tensor([1, 2048, 1, 256])]; + tensor var_1371_end_mask_0 = const()[name = string("op_1371_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1371 = slice_by_index(begin = var_1371_begin_0, end = var_1371_end_0, end_mask = var_1371_end_mask_0, x = cast_4)[name = string("op_1371")]; + tensor var_1383_axes_0 = const()[name = string("op_1383_axes_0"), val = tensor([-1])]; + tensor var_1383 = squeeze(axes = var_1383_axes_0, x = hidden_3)[name = string("op_1383")]; + tensor var_1385_axes_0 = const()[name = string("op_1385_axes_0"), val = tensor([-1])]; + tensor var_1385 = squeeze(axes = var_1385_axes_0, x = var_1383)[name = string("op_1385")]; + tensor hidden_states_25_axes_0 = const()[name = string("hidden_states_25_axes_0"), val = tensor([0])]; + tensor hidden_states_25 = expand_dims(axes = hidden_states_25_axes_0, x = var_1385)[name = string("hidden_states_25")]; + fp32 var_1391_promoted = const()[name = string("op_1391_promoted"), val = fp32(0x1p+1)]; + tensor var_1397 = pow(x = hidden_states_25, y = var_1391_promoted)[name = string("op_1397")]; + tensor variance_9_axes_0 = const()[name = string("variance_9_axes_0"), val = tensor([-1])]; + bool variance_9_keep_dims_0 = const()[name = string("variance_9_keep_dims_0"), val = bool(true)]; + tensor variance_9 = reduce_mean(axes = variance_9_axes_0, keep_dims = variance_9_keep_dims_0, x = var_1397)[name = string("variance_9")]; + fp32 var_1400 = const()[name = string("op_1400"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1401 = add(x = variance_9, y = var_1400)[name = string("op_1401")]; + fp32 var_1402_epsilon_0 = const()[name = string("op_1402_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1402 = rsqrt(epsilon = var_1402_epsilon_0, x = var_1401)[name = string("op_1402")]; + tensor hidden_states_29 = mul(x = hidden_states_25, y = var_1402)[name = string("hidden_states_29")]; + tensor const_11 = const()[name = string("const_11"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774237824)))]; + tensor input_11 = mul(x = const_11, y = hidden_states_29)[name = string("input_11")]; + tensor q_9 = linear(bias = linear_0_bias_0, weight = layers_1_self_attn_q_proj_weight, x = input_11)[name = string("linear_7")]; + tensor k_9 = linear(bias = linear_1_bias_0, weight = layers_1_self_attn_k_proj_weight, x = input_11)[name = string("linear_8")]; + tensor v_7 = linear(bias = linear_1_bias_0, weight = layers_1_self_attn_v_proj_weight, x = input_11)[name = string("linear_9")]; + tensor var_1419 = const()[name = string("op_1419"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_31 = reshape(shape = var_1419, x = q_9)[name = string("hidden_states_31")]; + tensor var_1425 = const()[name = string("op_1425"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_37 = reshape(shape = var_1425, x = k_9)[name = string("hidden_states_37")]; + tensor var_1431 = const()[name = string("op_1431"), val = tensor([1, 1, 8, 128])]; + tensor v_9 = reshape(shape = var_1431, x = v_7)[name = string("v_9")]; + fp32 var_1436_promoted = const()[name = string("op_1436_promoted"), val = fp32(0x1p+1)]; + tensor var_1442 = pow(x = hidden_states_31, y = var_1436_promoted)[name = string("op_1442")]; + tensor variance_11_axes_0 = const()[name = string("variance_11_axes_0"), val = tensor([-1])]; + bool variance_11_keep_dims_0 = const()[name = string("variance_11_keep_dims_0"), val = bool(true)]; + tensor variance_11 = reduce_mean(axes = variance_11_axes_0, keep_dims = variance_11_keep_dims_0, x = var_1442)[name = string("variance_11")]; + fp32 var_1445 = const()[name = string("op_1445"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1446 = add(x = variance_11, y = var_1445)[name = string("op_1446")]; + fp32 var_1447_epsilon_0 = const()[name = string("op_1447_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1447 = rsqrt(epsilon = var_1447_epsilon_0, x = var_1446)[name = string("op_1447")]; + tensor hidden_states_35 = mul(x = hidden_states_31, y = var_1447)[name = string("hidden_states_35")]; + tensor const_12 = const()[name = string("const_12"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774241984)))]; + tensor q_11 = mul(x = const_12, y = hidden_states_35)[name = string("q_11")]; + fp32 var_1454_promoted = const()[name = string("op_1454_promoted"), val = fp32(0x1p+1)]; + tensor var_1460 = pow(x = hidden_states_37, y = var_1454_promoted)[name = string("op_1460")]; + tensor variance_13_axes_0 = const()[name = string("variance_13_axes_0"), val = tensor([-1])]; + bool variance_13_keep_dims_0 = const()[name = string("variance_13_keep_dims_0"), val = bool(true)]; + tensor variance_13 = reduce_mean(axes = variance_13_axes_0, keep_dims = variance_13_keep_dims_0, x = var_1460)[name = string("variance_13")]; + fp32 var_1463 = const()[name = string("op_1463"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1464 = add(x = variance_13, y = var_1463)[name = string("op_1464")]; + fp32 var_1465_epsilon_0 = const()[name = string("op_1465_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1465 = rsqrt(epsilon = var_1465_epsilon_0, x = var_1464)[name = string("op_1465")]; + tensor hidden_states_41 = mul(x = hidden_states_37, y = var_1465)[name = string("hidden_states_41")]; + tensor const_13 = const()[name = string("const_13"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774242560)))]; + tensor k_11 = mul(x = const_13, y = hidden_states_41)[name = string("k_11")]; + tensor q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = q_11)[name = string("transpose_107")]; + tensor var_1482 = mul(x = q_13, y = cos_3)[name = string("op_1482")]; + tensor x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_5 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_13)[name = string("x1_5")]; + tensor x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_5 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_13)[name = string("x2_5")]; + fp32 const_16_promoted = const()[name = string("const_16_promoted"), val = fp32(-0x1p+0)]; + tensor var_1503 = mul(x = x2_5, y = const_16_promoted)[name = string("op_1503")]; + int32 var_1505 = const()[name = string("op_1505"), val = int32(-1)]; + bool var_1506_interleave_0 = const()[name = string("op_1506_interleave_0"), val = bool(false)]; + tensor var_1506 = concat(axis = var_1505, interleave = var_1506_interleave_0, values = (var_1503, x1_5))[name = string("op_1506")]; + tensor var_1507 = mul(x = var_1506, y = sin_3)[name = string("op_1507")]; + tensor q_15 = add(x = var_1482, y = var_1507)[name = string("q_15")]; + tensor k_13 = transpose(perm = k_13_perm_0, x = k_11)[name = string("transpose_106")]; + tensor var_1510 = mul(x = k_13, y = cos_3)[name = string("op_1510")]; + tensor x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_7 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_13)[name = string("x1_7")]; + tensor x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_7 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_13)[name = string("x2_7")]; + fp32 const_19_promoted = const()[name = string("const_19_promoted"), val = fp32(-0x1p+0)]; + tensor var_1531 = mul(x = x2_7, y = const_19_promoted)[name = string("op_1531")]; + int32 var_1533 = const()[name = string("op_1533"), val = int32(-1)]; + bool var_1534_interleave_0 = const()[name = string("op_1534_interleave_0"), val = bool(false)]; + tensor var_1534 = concat(axis = var_1533, interleave = var_1534_interleave_0, values = (var_1531, x1_7))[name = string("op_1534")]; + tensor var_1535 = mul(x = var_1534, y = sin_3)[name = string("op_1535")]; + tensor k_15 = add(x = var_1510, y = var_1535)[name = string("k_15")]; + tensor var_1542 = const()[name = string("op_1542"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_3 = reshape(shape = var_1542, x = k_15)[name = string("nk_flat_3")]; + tensor var_1548 = const()[name = string("op_1548"), val = tensor([1, 1024, 1, 1])]; + tensor v_11 = transpose(perm = v_11_perm_0, x = v_9)[name = string("transpose_105")]; + tensor nv_flat_3 = reshape(shape = var_1548, x = v_11)[name = string("nv_flat_3")]; + tensor var_1557 = mul(x = var_1351, y = var_1194)[name = string("op_1557")]; + tensor var_1558 = mul(x = nk_flat_3, y = update_mask_1)[name = string("op_1558")]; + tensor key_cache_9 = add(x = var_1557, y = var_1558)[name = string("key_cache_9")]; + tensor var_1564 = mul(x = var_1371, y = var_1194)[name = string("op_1564")]; + tensor var_1565 = mul(x = nv_flat_3, y = update_mask_1)[name = string("op_1565")]; + tensor value_cache_9 = add(x = var_1564, y = var_1565)[name = string("value_cache_9")]; + tensor kc_7_axes_0 = const()[name = string("kc_7_axes_0"), val = tensor([2])]; + tensor kc_7 = squeeze(axes = kc_7_axes_0, x = key_cache_9)[name = string("kc_7")]; + tensor var_1574 = const()[name = string("op_1574"), val = tensor([1, 8, 128, 256])]; + tensor kc_9 = reshape(shape = var_1574, x = kc_7)[name = string("kc_9")]; + tensor vc_7_axes_0 = const()[name = string("vc_7_axes_0"), val = tensor([2])]; + tensor vc_7 = squeeze(axes = vc_7_axes_0, x = value_cache_9)[name = string("vc_7")]; + tensor var_1582 = const()[name = string("op_1582"), val = tensor([1, 8, 128, 256])]; + tensor vc_9 = reshape(shape = var_1582, x = vc_7)[name = string("vc_9")]; + tensor var_1585_axes_0 = const()[name = string("op_1585_axes_0"), val = tensor([2])]; + tensor var_1585 = expand_dims(axes = var_1585_axes_0, x = kc_9)[name = string("op_1585")]; + tensor var_1593_reps_0 = const()[name = string("op_1593_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_1593 = tile(reps = var_1593_reps_0, x = var_1585)[name = string("op_1593")]; + tensor var_1598 = const()[name = string("op_1598"), val = tensor([1, 16, 128, 256])]; + tensor kc_11 = reshape(shape = var_1598, x = var_1593)[name = string("kc_11")]; + tensor var_1601_axes_0 = const()[name = string("op_1601_axes_0"), val = tensor([2])]; + tensor var_1601 = expand_dims(axes = var_1601_axes_0, x = vc_9)[name = string("op_1601")]; + tensor var_1609_reps_0 = const()[name = string("op_1609_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_1609 = tile(reps = var_1609_reps_0, x = var_1601)[name = string("op_1609")]; + tensor var_1614 = const()[name = string("op_1614"), val = tensor([1, 16, 128, 256])]; + tensor vc_11 = reshape(shape = var_1614, x = var_1609)[name = string("vc_11")]; + bool var_1616_transpose_x_0 = const()[name = string("op_1616_transpose_x_0"), val = bool(false)]; + bool var_1616_transpose_y_0 = const()[name = string("op_1616_transpose_y_0"), val = bool(false)]; + tensor var_1616 = matmul(transpose_x = var_1616_transpose_x_0, transpose_y = var_1616_transpose_y_0, x = q_15, y = kc_11)[name = string("op_1616")]; + fp32 _inversed_attn_weights_9_y_0 = const()[name = string("_inversed_attn_weights_9_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_9 = mul(x = var_1616, y = _inversed_attn_weights_9_y_0)[name = string("_inversed_attn_weights_9")]; + tensor attn_weights_11 = add(x = _inversed_attn_weights_9, y = mask_1)[name = string("attn_weights_11")]; + int32 var_1630 = const()[name = string("op_1630"), val = int32(-1)]; + tensor attn_weights_15 = softmax(axis = var_1630, x = attn_weights_11)[name = string("attn_weights_15")]; + bool attn_output_5_transpose_x_1 = const()[name = string("attn_output_5_transpose_x_1"), val = bool(false)]; + bool attn_output_5_transpose_y_1 = const()[name = string("attn_output_5_transpose_y_1"), val = bool(true)]; + tensor attn_output_5 = matmul(transpose_x = attn_output_5_transpose_x_1, transpose_y = attn_output_5_transpose_y_1, x = attn_weights_15, y = vc_11)[name = string("attn_output_5")]; + tensor var_1639_perm_0 = const()[name = string("op_1639_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1643 = const()[name = string("op_1643"), val = tensor([1, 1, -1])]; + tensor var_1639 = transpose(perm = var_1639_perm_0, x = attn_output_5)[name = string("transpose_104")]; + tensor input_13 = reshape(shape = var_1643, x = var_1639)[name = string("input_13")]; + tensor attn_output_7 = linear(bias = linear_1_bias_0, weight = layers_1_self_attn_o_proj_weight, x = input_13)[name = string("linear_10")]; + tensor var_1649_axes_0 = const()[name = string("op_1649_axes_0"), val = tensor([0])]; + tensor var_1649 = squeeze(axes = var_1649_axes_0, x = attn_output_7)[name = string("op_1649")]; + tensor var_1651_axes_0 = const()[name = string("op_1651_axes_0"), val = tensor([0])]; + tensor var_1651 = squeeze(axes = var_1651_axes_0, x = var_1649)[name = string("op_1651")]; + tensor var_1653_axes_0 = const()[name = string("op_1653_axes_0"), val = tensor([-1])]; + tensor var_1653 = expand_dims(axes = var_1653_axes_0, x = var_1651)[name = string("op_1653")]; + tensor attn_4d_3_axes_0 = const()[name = string("attn_4d_3_axes_0"), val = tensor([-1])]; + tensor attn_4d_3 = expand_dims(axes = attn_4d_3_axes_0, x = var_1653)[name = string("attn_4d_3")]; + tensor hidden_5 = add(x = hidden_3, y = attn_4d_3)[name = string("hidden_5")]; + tensor var_1659_axes_0 = const()[name = string("op_1659_axes_0"), val = tensor([-1])]; + tensor var_1659 = squeeze(axes = var_1659_axes_0, x = hidden_5)[name = string("op_1659")]; + tensor var_1661_axes_0 = const()[name = string("op_1661_axes_0"), val = tensor([-1])]; + tensor var_1661 = squeeze(axes = var_1661_axes_0, x = var_1659)[name = string("op_1661")]; + tensor hidden_states_43_axes_0 = const()[name = string("hidden_states_43_axes_0"), val = tensor([0])]; + tensor hidden_states_43 = expand_dims(axes = hidden_states_43_axes_0, x = var_1661)[name = string("hidden_states_43")]; + fp32 var_1667_promoted = const()[name = string("op_1667_promoted"), val = fp32(0x1p+1)]; + tensor var_1673 = pow(x = hidden_states_43, y = var_1667_promoted)[name = string("op_1673")]; + tensor variance_15_axes_0 = const()[name = string("variance_15_axes_0"), val = tensor([-1])]; + bool variance_15_keep_dims_0 = const()[name = string("variance_15_keep_dims_0"), val = bool(true)]; + tensor variance_15 = reduce_mean(axes = variance_15_axes_0, keep_dims = variance_15_keep_dims_0, x = var_1673)[name = string("variance_15")]; + fp32 var_1676 = const()[name = string("op_1676"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1677 = add(x = variance_15, y = var_1676)[name = string("op_1677")]; + fp32 var_1678_epsilon_0 = const()[name = string("op_1678_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1678 = rsqrt(epsilon = var_1678_epsilon_0, x = var_1677)[name = string("op_1678")]; + tensor hidden_states_47 = mul(x = hidden_states_43, y = var_1678)[name = string("hidden_states_47")]; + tensor const_20 = const()[name = string("const_20"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774243136)))]; + tensor input_15 = mul(x = const_20, y = hidden_states_47)[name = string("input_15")]; + tensor input_17 = linear(bias = linear_4_bias_0, weight = layers_1_mlp_gate_proj_weight, x = input_15)[name = string("linear_11")]; + tensor var_1688 = silu(x = input_17)[name = string("op_1688")]; + tensor var_1690 = linear(bias = linear_4_bias_0, weight = layers_1_mlp_up_proj_weight, x = input_15)[name = string("linear_12")]; + tensor input_19 = mul(x = var_1688, y = var_1690)[name = string("input_19")]; + tensor mlp_out_3 = linear(bias = linear_1_bias_0, weight = layers_1_mlp_down_proj_weight, x = input_19)[name = string("linear_13")]; + tensor var_1695_axes_0 = const()[name = string("op_1695_axes_0"), val = tensor([0])]; + tensor var_1695 = squeeze(axes = var_1695_axes_0, x = mlp_out_3)[name = string("op_1695")]; + tensor var_1697_axes_0 = const()[name = string("op_1697_axes_0"), val = tensor([0])]; + tensor var_1697 = squeeze(axes = var_1697_axes_0, x = var_1695)[name = string("op_1697")]; + tensor var_1699_axes_0 = const()[name = string("op_1699_axes_0"), val = tensor([-1])]; + tensor var_1699 = expand_dims(axes = var_1699_axes_0, x = var_1697)[name = string("op_1699")]; + tensor mlp_4d_3_axes_0 = const()[name = string("mlp_4d_3_axes_0"), val = tensor([-1])]; + tensor mlp_4d_3 = expand_dims(axes = mlp_4d_3_axes_0, x = var_1699)[name = string("mlp_4d_3")]; + tensor hidden_7 = add(x = hidden_5, y = mlp_4d_3)[name = string("hidden_7")]; + tensor var_1713_begin_0 = const()[name = string("op_1713_begin_0"), val = tensor([0, 2048, 0, 0])]; + tensor var_1713_end_0 = const()[name = string("op_1713_end_0"), val = tensor([1, 3072, 1, 256])]; + tensor var_1713_end_mask_0 = const()[name = string("op_1713_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1713 = slice_by_index(begin = var_1713_begin_0, end = var_1713_end_0, end_mask = var_1713_end_mask_0, x = cast_3)[name = string("op_1713")]; + tensor var_1733_begin_0 = const()[name = string("op_1733_begin_0"), val = tensor([0, 2048, 0, 0])]; + tensor var_1733_end_0 = const()[name = string("op_1733_end_0"), val = tensor([1, 3072, 1, 256])]; + tensor var_1733_end_mask_0 = const()[name = string("op_1733_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1733 = slice_by_index(begin = var_1733_begin_0, end = var_1733_end_0, end_mask = var_1733_end_mask_0, x = cast_4)[name = string("op_1733")]; + tensor var_1745_axes_0 = const()[name = string("op_1745_axes_0"), val = tensor([-1])]; + tensor var_1745 = squeeze(axes = var_1745_axes_0, x = hidden_7)[name = string("op_1745")]; + tensor var_1747_axes_0 = const()[name = string("op_1747_axes_0"), val = tensor([-1])]; + tensor var_1747 = squeeze(axes = var_1747_axes_0, x = var_1745)[name = string("op_1747")]; + tensor hidden_states_49_axes_0 = const()[name = string("hidden_states_49_axes_0"), val = tensor([0])]; + tensor hidden_states_49 = expand_dims(axes = hidden_states_49_axes_0, x = var_1747)[name = string("hidden_states_49")]; + fp32 var_1753_promoted = const()[name = string("op_1753_promoted"), val = fp32(0x1p+1)]; + tensor var_1759 = pow(x = hidden_states_49, y = var_1753_promoted)[name = string("op_1759")]; + tensor variance_17_axes_0 = const()[name = string("variance_17_axes_0"), val = tensor([-1])]; + bool variance_17_keep_dims_0 = const()[name = string("variance_17_keep_dims_0"), val = bool(true)]; + tensor variance_17 = reduce_mean(axes = variance_17_axes_0, keep_dims = variance_17_keep_dims_0, x = var_1759)[name = string("variance_17")]; + fp32 var_1762 = const()[name = string("op_1762"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1763 = add(x = variance_17, y = var_1762)[name = string("op_1763")]; + fp32 var_1764_epsilon_0 = const()[name = string("op_1764_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1764 = rsqrt(epsilon = var_1764_epsilon_0, x = var_1763)[name = string("op_1764")]; + tensor hidden_states_53 = mul(x = hidden_states_49, y = var_1764)[name = string("hidden_states_53")]; + tensor const_21 = const()[name = string("const_21"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774247296)))]; + tensor input_21 = mul(x = const_21, y = hidden_states_53)[name = string("input_21")]; + tensor q_17 = linear(bias = linear_0_bias_0, weight = layers_2_self_attn_q_proj_weight, x = input_21)[name = string("linear_14")]; + tensor k_17 = linear(bias = linear_1_bias_0, weight = layers_2_self_attn_k_proj_weight, x = input_21)[name = string("linear_15")]; + tensor v_13 = linear(bias = linear_1_bias_0, weight = layers_2_self_attn_v_proj_weight, x = input_21)[name = string("linear_16")]; + tensor var_1781 = const()[name = string("op_1781"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_55 = reshape(shape = var_1781, x = q_17)[name = string("hidden_states_55")]; + tensor var_1787 = const()[name = string("op_1787"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_61 = reshape(shape = var_1787, x = k_17)[name = string("hidden_states_61")]; + tensor var_1793 = const()[name = string("op_1793"), val = tensor([1, 1, 8, 128])]; + tensor v_15 = reshape(shape = var_1793, x = v_13)[name = string("v_15")]; + fp32 var_1798_promoted = const()[name = string("op_1798_promoted"), val = fp32(0x1p+1)]; + tensor var_1804 = pow(x = hidden_states_55, y = var_1798_promoted)[name = string("op_1804")]; + tensor variance_19_axes_0 = const()[name = string("variance_19_axes_0"), val = tensor([-1])]; + bool variance_19_keep_dims_0 = const()[name = string("variance_19_keep_dims_0"), val = bool(true)]; + tensor variance_19 = reduce_mean(axes = variance_19_axes_0, keep_dims = variance_19_keep_dims_0, x = var_1804)[name = string("variance_19")]; + fp32 var_1807 = const()[name = string("op_1807"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1808 = add(x = variance_19, y = var_1807)[name = string("op_1808")]; + fp32 var_1809_epsilon_0 = const()[name = string("op_1809_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1809 = rsqrt(epsilon = var_1809_epsilon_0, x = var_1808)[name = string("op_1809")]; + tensor hidden_states_59 = mul(x = hidden_states_55, y = var_1809)[name = string("hidden_states_59")]; + tensor const_22 = const()[name = string("const_22"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774251456)))]; + tensor q_19 = mul(x = const_22, y = hidden_states_59)[name = string("q_19")]; + fp32 var_1816_promoted = const()[name = string("op_1816_promoted"), val = fp32(0x1p+1)]; + tensor var_1822 = pow(x = hidden_states_61, y = var_1816_promoted)[name = string("op_1822")]; + tensor variance_21_axes_0 = const()[name = string("variance_21_axes_0"), val = tensor([-1])]; + bool variance_21_keep_dims_0 = const()[name = string("variance_21_keep_dims_0"), val = bool(true)]; + tensor variance_21 = reduce_mean(axes = variance_21_axes_0, keep_dims = variance_21_keep_dims_0, x = var_1822)[name = string("variance_21")]; + fp32 var_1825 = const()[name = string("op_1825"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_1826 = add(x = variance_21, y = var_1825)[name = string("op_1826")]; + fp32 var_1827_epsilon_0 = const()[name = string("op_1827_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_1827 = rsqrt(epsilon = var_1827_epsilon_0, x = var_1826)[name = string("op_1827")]; + tensor hidden_states_65 = mul(x = hidden_states_61, y = var_1827)[name = string("hidden_states_65")]; + tensor const_23 = const()[name = string("const_23"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774252032)))]; + tensor k_19 = mul(x = const_23, y = hidden_states_65)[name = string("k_19")]; + tensor q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = q_19)[name = string("transpose_103")]; + tensor var_1844 = mul(x = q_21, y = cos_3)[name = string("op_1844")]; + tensor x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_9 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_21)[name = string("x1_9")]; + tensor x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_9 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_21)[name = string("x2_9")]; + fp32 const_26_promoted = const()[name = string("const_26_promoted"), val = fp32(-0x1p+0)]; + tensor var_1865 = mul(x = x2_9, y = const_26_promoted)[name = string("op_1865")]; + int32 var_1867 = const()[name = string("op_1867"), val = int32(-1)]; + bool var_1868_interleave_0 = const()[name = string("op_1868_interleave_0"), val = bool(false)]; + tensor var_1868 = concat(axis = var_1867, interleave = var_1868_interleave_0, values = (var_1865, x1_9))[name = string("op_1868")]; + tensor var_1869 = mul(x = var_1868, y = sin_3)[name = string("op_1869")]; + tensor q_23 = add(x = var_1844, y = var_1869)[name = string("q_23")]; + tensor k_21 = transpose(perm = k_21_perm_0, x = k_19)[name = string("transpose_102")]; + tensor var_1872 = mul(x = k_21, y = cos_3)[name = string("op_1872")]; + tensor x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_11 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_21)[name = string("x1_11")]; + tensor x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_11 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_21)[name = string("x2_11")]; + fp32 const_29_promoted = const()[name = string("const_29_promoted"), val = fp32(-0x1p+0)]; + tensor var_1893 = mul(x = x2_11, y = const_29_promoted)[name = string("op_1893")]; + int32 var_1895 = const()[name = string("op_1895"), val = int32(-1)]; + bool var_1896_interleave_0 = const()[name = string("op_1896_interleave_0"), val = bool(false)]; + tensor var_1896 = concat(axis = var_1895, interleave = var_1896_interleave_0, values = (var_1893, x1_11))[name = string("op_1896")]; + tensor var_1897 = mul(x = var_1896, y = sin_3)[name = string("op_1897")]; + tensor k_23 = add(x = var_1872, y = var_1897)[name = string("k_23")]; + tensor var_1904 = const()[name = string("op_1904"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_5 = reshape(shape = var_1904, x = k_23)[name = string("nk_flat_5")]; + tensor var_1910 = const()[name = string("op_1910"), val = tensor([1, 1024, 1, 1])]; + tensor v_17 = transpose(perm = v_17_perm_0, x = v_15)[name = string("transpose_101")]; + tensor nv_flat_5 = reshape(shape = var_1910, x = v_17)[name = string("nv_flat_5")]; + tensor var_1919 = mul(x = var_1713, y = var_1194)[name = string("op_1919")]; + tensor var_1920 = mul(x = nk_flat_5, y = update_mask_1)[name = string("op_1920")]; + tensor key_cache_13 = add(x = var_1919, y = var_1920)[name = string("key_cache_13")]; + tensor var_1926 = mul(x = var_1733, y = var_1194)[name = string("op_1926")]; + tensor var_1927 = mul(x = nv_flat_5, y = update_mask_1)[name = string("op_1927")]; + tensor value_cache_13 = add(x = var_1926, y = var_1927)[name = string("value_cache_13")]; + tensor kc_13_axes_0 = const()[name = string("kc_13_axes_0"), val = tensor([2])]; + tensor kc_13 = squeeze(axes = kc_13_axes_0, x = key_cache_13)[name = string("kc_13")]; + tensor var_1936 = const()[name = string("op_1936"), val = tensor([1, 8, 128, 256])]; + tensor kc_15 = reshape(shape = var_1936, x = kc_13)[name = string("kc_15")]; + tensor vc_13_axes_0 = const()[name = string("vc_13_axes_0"), val = tensor([2])]; + tensor vc_13 = squeeze(axes = vc_13_axes_0, x = value_cache_13)[name = string("vc_13")]; + tensor var_1944 = const()[name = string("op_1944"), val = tensor([1, 8, 128, 256])]; + tensor vc_15 = reshape(shape = var_1944, x = vc_13)[name = string("vc_15")]; + tensor var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor([2])]; + tensor var_1947 = expand_dims(axes = var_1947_axes_0, x = kc_15)[name = string("op_1947")]; + tensor var_1955_reps_0 = const()[name = string("op_1955_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_1955 = tile(reps = var_1955_reps_0, x = var_1947)[name = string("op_1955")]; + tensor var_1960 = const()[name = string("op_1960"), val = tensor([1, 16, 128, 256])]; + tensor kc_17 = reshape(shape = var_1960, x = var_1955)[name = string("kc_17")]; + tensor var_1963_axes_0 = const()[name = string("op_1963_axes_0"), val = tensor([2])]; + tensor var_1963 = expand_dims(axes = var_1963_axes_0, x = vc_15)[name = string("op_1963")]; + tensor var_1971_reps_0 = const()[name = string("op_1971_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_1971 = tile(reps = var_1971_reps_0, x = var_1963)[name = string("op_1971")]; + tensor var_1976 = const()[name = string("op_1976"), val = tensor([1, 16, 128, 256])]; + tensor vc_17 = reshape(shape = var_1976, x = var_1971)[name = string("vc_17")]; + bool var_1978_transpose_x_0 = const()[name = string("op_1978_transpose_x_0"), val = bool(false)]; + bool var_1978_transpose_y_0 = const()[name = string("op_1978_transpose_y_0"), val = bool(false)]; + tensor var_1978 = matmul(transpose_x = var_1978_transpose_x_0, transpose_y = var_1978_transpose_y_0, x = q_23, y = kc_17)[name = string("op_1978")]; + fp32 _inversed_attn_weights_17_y_0 = const()[name = string("_inversed_attn_weights_17_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_17 = mul(x = var_1978, y = _inversed_attn_weights_17_y_0)[name = string("_inversed_attn_weights_17")]; + tensor attn_weights_19 = add(x = _inversed_attn_weights_17, y = mask_1)[name = string("attn_weights_19")]; + int32 var_1992 = const()[name = string("op_1992"), val = int32(-1)]; + tensor attn_weights_23 = softmax(axis = var_1992, x = attn_weights_19)[name = string("attn_weights_23")]; + bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(false)]; + bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(true)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_23, y = vc_17)[name = string("attn_output_9")]; + tensor var_2001_perm_0 = const()[name = string("op_2001_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2005 = const()[name = string("op_2005"), val = tensor([1, 1, -1])]; + tensor var_2001 = transpose(perm = var_2001_perm_0, x = attn_output_9)[name = string("transpose_100")]; + tensor input_23 = reshape(shape = var_2005, x = var_2001)[name = string("input_23")]; + tensor attn_output_11 = linear(bias = linear_1_bias_0, weight = layers_2_self_attn_o_proj_weight, x = input_23)[name = string("linear_17")]; + tensor var_2011_axes_0 = const()[name = string("op_2011_axes_0"), val = tensor([0])]; + tensor var_2011 = squeeze(axes = var_2011_axes_0, x = attn_output_11)[name = string("op_2011")]; + tensor var_2013_axes_0 = const()[name = string("op_2013_axes_0"), val = tensor([0])]; + tensor var_2013 = squeeze(axes = var_2013_axes_0, x = var_2011)[name = string("op_2013")]; + tensor var_2015_axes_0 = const()[name = string("op_2015_axes_0"), val = tensor([-1])]; + tensor var_2015 = expand_dims(axes = var_2015_axes_0, x = var_2013)[name = string("op_2015")]; + tensor attn_4d_5_axes_0 = const()[name = string("attn_4d_5_axes_0"), val = tensor([-1])]; + tensor attn_4d_5 = expand_dims(axes = attn_4d_5_axes_0, x = var_2015)[name = string("attn_4d_5")]; + tensor hidden_9 = add(x = hidden_7, y = attn_4d_5)[name = string("hidden_9")]; + tensor var_2021_axes_0 = const()[name = string("op_2021_axes_0"), val = tensor([-1])]; + tensor var_2021 = squeeze(axes = var_2021_axes_0, x = hidden_9)[name = string("op_2021")]; + tensor var_2023_axes_0 = const()[name = string("op_2023_axes_0"), val = tensor([-1])]; + tensor var_2023 = squeeze(axes = var_2023_axes_0, x = var_2021)[name = string("op_2023")]; + tensor hidden_states_67_axes_0 = const()[name = string("hidden_states_67_axes_0"), val = tensor([0])]; + tensor hidden_states_67 = expand_dims(axes = hidden_states_67_axes_0, x = var_2023)[name = string("hidden_states_67")]; + fp32 var_2029_promoted = const()[name = string("op_2029_promoted"), val = fp32(0x1p+1)]; + tensor var_2035 = pow(x = hidden_states_67, y = var_2029_promoted)[name = string("op_2035")]; + tensor variance_23_axes_0 = const()[name = string("variance_23_axes_0"), val = tensor([-1])]; + bool variance_23_keep_dims_0 = const()[name = string("variance_23_keep_dims_0"), val = bool(true)]; + tensor variance_23 = reduce_mean(axes = variance_23_axes_0, keep_dims = variance_23_keep_dims_0, x = var_2035)[name = string("variance_23")]; + fp32 var_2038 = const()[name = string("op_2038"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2039 = add(x = variance_23, y = var_2038)[name = string("op_2039")]; + fp32 var_2040_epsilon_0 = const()[name = string("op_2040_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2040 = rsqrt(epsilon = var_2040_epsilon_0, x = var_2039)[name = string("op_2040")]; + tensor hidden_states_71 = mul(x = hidden_states_67, y = var_2040)[name = string("hidden_states_71")]; + tensor const_30 = const()[name = string("const_30"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774252608)))]; + tensor input_25 = mul(x = const_30, y = hidden_states_71)[name = string("input_25")]; + tensor input_27 = linear(bias = linear_4_bias_0, weight = layers_2_mlp_gate_proj_weight, x = input_25)[name = string("linear_18")]; + tensor var_2050 = silu(x = input_27)[name = string("op_2050")]; + tensor var_2052 = linear(bias = linear_4_bias_0, weight = layers_2_mlp_up_proj_weight, x = input_25)[name = string("linear_19")]; + tensor input_29 = mul(x = var_2050, y = var_2052)[name = string("input_29")]; + tensor mlp_out_5 = linear(bias = linear_1_bias_0, weight = layers_2_mlp_down_proj_weight, x = input_29)[name = string("linear_20")]; + tensor var_2057_axes_0 = const()[name = string("op_2057_axes_0"), val = tensor([0])]; + tensor var_2057 = squeeze(axes = var_2057_axes_0, x = mlp_out_5)[name = string("op_2057")]; + tensor var_2059_axes_0 = const()[name = string("op_2059_axes_0"), val = tensor([0])]; + tensor var_2059 = squeeze(axes = var_2059_axes_0, x = var_2057)[name = string("op_2059")]; + tensor var_2061_axes_0 = const()[name = string("op_2061_axes_0"), val = tensor([-1])]; + tensor var_2061 = expand_dims(axes = var_2061_axes_0, x = var_2059)[name = string("op_2061")]; + tensor mlp_4d_5_axes_0 = const()[name = string("mlp_4d_5_axes_0"), val = tensor([-1])]; + tensor mlp_4d_5 = expand_dims(axes = mlp_4d_5_axes_0, x = var_2061)[name = string("mlp_4d_5")]; + tensor hidden_11 = add(x = hidden_9, y = mlp_4d_5)[name = string("hidden_11")]; + tensor var_2075_begin_0 = const()[name = string("op_2075_begin_0"), val = tensor([0, 3072, 0, 0])]; + tensor var_2075_end_0 = const()[name = string("op_2075_end_0"), val = tensor([1, 4096, 1, 256])]; + tensor var_2075_end_mask_0 = const()[name = string("op_2075_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2075 = slice_by_index(begin = var_2075_begin_0, end = var_2075_end_0, end_mask = var_2075_end_mask_0, x = cast_3)[name = string("op_2075")]; + tensor var_2095_begin_0 = const()[name = string("op_2095_begin_0"), val = tensor([0, 3072, 0, 0])]; + tensor var_2095_end_0 = const()[name = string("op_2095_end_0"), val = tensor([1, 4096, 1, 256])]; + tensor var_2095_end_mask_0 = const()[name = string("op_2095_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2095 = slice_by_index(begin = var_2095_begin_0, end = var_2095_end_0, end_mask = var_2095_end_mask_0, x = cast_4)[name = string("op_2095")]; + tensor var_2107_axes_0 = const()[name = string("op_2107_axes_0"), val = tensor([-1])]; + tensor var_2107 = squeeze(axes = var_2107_axes_0, x = hidden_11)[name = string("op_2107")]; + tensor var_2109_axes_0 = const()[name = string("op_2109_axes_0"), val = tensor([-1])]; + tensor var_2109 = squeeze(axes = var_2109_axes_0, x = var_2107)[name = string("op_2109")]; + tensor hidden_states_73_axes_0 = const()[name = string("hidden_states_73_axes_0"), val = tensor([0])]; + tensor hidden_states_73 = expand_dims(axes = hidden_states_73_axes_0, x = var_2109)[name = string("hidden_states_73")]; + fp32 var_2115_promoted = const()[name = string("op_2115_promoted"), val = fp32(0x1p+1)]; + tensor var_2121 = pow(x = hidden_states_73, y = var_2115_promoted)[name = string("op_2121")]; + tensor variance_25_axes_0 = const()[name = string("variance_25_axes_0"), val = tensor([-1])]; + bool variance_25_keep_dims_0 = const()[name = string("variance_25_keep_dims_0"), val = bool(true)]; + tensor variance_25 = reduce_mean(axes = variance_25_axes_0, keep_dims = variance_25_keep_dims_0, x = var_2121)[name = string("variance_25")]; + fp32 var_2124 = const()[name = string("op_2124"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2125 = add(x = variance_25, y = var_2124)[name = string("op_2125")]; + fp32 var_2126_epsilon_0 = const()[name = string("op_2126_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2126 = rsqrt(epsilon = var_2126_epsilon_0, x = var_2125)[name = string("op_2126")]; + tensor hidden_states_77 = mul(x = hidden_states_73, y = var_2126)[name = string("hidden_states_77")]; + tensor const_31 = const()[name = string("const_31"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774256768)))]; + tensor input_31 = mul(x = const_31, y = hidden_states_77)[name = string("input_31")]; + tensor q_25 = linear(bias = linear_0_bias_0, weight = layers_3_self_attn_q_proj_weight, x = input_31)[name = string("linear_21")]; + tensor k_25 = linear(bias = linear_1_bias_0, weight = layers_3_self_attn_k_proj_weight, x = input_31)[name = string("linear_22")]; + tensor v_19 = linear(bias = linear_1_bias_0, weight = layers_3_self_attn_v_proj_weight, x = input_31)[name = string("linear_23")]; + tensor var_2143 = const()[name = string("op_2143"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_79 = reshape(shape = var_2143, x = q_25)[name = string("hidden_states_79")]; + tensor var_2149 = const()[name = string("op_2149"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_85 = reshape(shape = var_2149, x = k_25)[name = string("hidden_states_85")]; + tensor var_2155 = const()[name = string("op_2155"), val = tensor([1, 1, 8, 128])]; + tensor v_21 = reshape(shape = var_2155, x = v_19)[name = string("v_21")]; + fp32 var_2160_promoted = const()[name = string("op_2160_promoted"), val = fp32(0x1p+1)]; + tensor var_2166 = pow(x = hidden_states_79, y = var_2160_promoted)[name = string("op_2166")]; + tensor variance_27_axes_0 = const()[name = string("variance_27_axes_0"), val = tensor([-1])]; + bool variance_27_keep_dims_0 = const()[name = string("variance_27_keep_dims_0"), val = bool(true)]; + tensor variance_27 = reduce_mean(axes = variance_27_axes_0, keep_dims = variance_27_keep_dims_0, x = var_2166)[name = string("variance_27")]; + fp32 var_2169 = const()[name = string("op_2169"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2170 = add(x = variance_27, y = var_2169)[name = string("op_2170")]; + fp32 var_2171_epsilon_0 = const()[name = string("op_2171_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2171 = rsqrt(epsilon = var_2171_epsilon_0, x = var_2170)[name = string("op_2171")]; + tensor hidden_states_83 = mul(x = hidden_states_79, y = var_2171)[name = string("hidden_states_83")]; + tensor const_32 = const()[name = string("const_32"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774260928)))]; + tensor q_27 = mul(x = const_32, y = hidden_states_83)[name = string("q_27")]; + fp32 var_2178_promoted = const()[name = string("op_2178_promoted"), val = fp32(0x1p+1)]; + tensor var_2184 = pow(x = hidden_states_85, y = var_2178_promoted)[name = string("op_2184")]; + tensor variance_29_axes_0 = const()[name = string("variance_29_axes_0"), val = tensor([-1])]; + bool variance_29_keep_dims_0 = const()[name = string("variance_29_keep_dims_0"), val = bool(true)]; + tensor variance_29 = reduce_mean(axes = variance_29_axes_0, keep_dims = variance_29_keep_dims_0, x = var_2184)[name = string("variance_29")]; + fp32 var_2187 = const()[name = string("op_2187"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2188 = add(x = variance_29, y = var_2187)[name = string("op_2188")]; + fp32 var_2189_epsilon_0 = const()[name = string("op_2189_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2189 = rsqrt(epsilon = var_2189_epsilon_0, x = var_2188)[name = string("op_2189")]; + tensor hidden_states_89 = mul(x = hidden_states_85, y = var_2189)[name = string("hidden_states_89")]; + tensor const_33 = const()[name = string("const_33"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774261504)))]; + tensor k_27 = mul(x = const_33, y = hidden_states_89)[name = string("k_27")]; + tensor q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_29 = transpose(perm = q_29_perm_0, x = q_27)[name = string("transpose_99")]; + tensor var_2206 = mul(x = q_29, y = cos_3)[name = string("op_2206")]; + tensor x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_13 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_29)[name = string("x1_13")]; + tensor x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_13 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_29)[name = string("x2_13")]; + fp32 const_36_promoted = const()[name = string("const_36_promoted"), val = fp32(-0x1p+0)]; + tensor var_2227 = mul(x = x2_13, y = const_36_promoted)[name = string("op_2227")]; + int32 var_2229 = const()[name = string("op_2229"), val = int32(-1)]; + bool var_2230_interleave_0 = const()[name = string("op_2230_interleave_0"), val = bool(false)]; + tensor var_2230 = concat(axis = var_2229, interleave = var_2230_interleave_0, values = (var_2227, x1_13))[name = string("op_2230")]; + tensor var_2231 = mul(x = var_2230, y = sin_3)[name = string("op_2231")]; + tensor q_31 = add(x = var_2206, y = var_2231)[name = string("q_31")]; + tensor k_29 = transpose(perm = k_29_perm_0, x = k_27)[name = string("transpose_98")]; + tensor var_2234 = mul(x = k_29, y = cos_3)[name = string("op_2234")]; + tensor x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_15 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_29)[name = string("x1_15")]; + tensor x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_15 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_29)[name = string("x2_15")]; + fp32 const_39_promoted = const()[name = string("const_39_promoted"), val = fp32(-0x1p+0)]; + tensor var_2255 = mul(x = x2_15, y = const_39_promoted)[name = string("op_2255")]; + int32 var_2257 = const()[name = string("op_2257"), val = int32(-1)]; + bool var_2258_interleave_0 = const()[name = string("op_2258_interleave_0"), val = bool(false)]; + tensor var_2258 = concat(axis = var_2257, interleave = var_2258_interleave_0, values = (var_2255, x1_15))[name = string("op_2258")]; + tensor var_2259 = mul(x = var_2258, y = sin_3)[name = string("op_2259")]; + tensor k_31 = add(x = var_2234, y = var_2259)[name = string("k_31")]; + tensor var_2266 = const()[name = string("op_2266"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_7 = reshape(shape = var_2266, x = k_31)[name = string("nk_flat_7")]; + tensor var_2272 = const()[name = string("op_2272"), val = tensor([1, 1024, 1, 1])]; + tensor v_23 = transpose(perm = v_23_perm_0, x = v_21)[name = string("transpose_97")]; + tensor nv_flat_7 = reshape(shape = var_2272, x = v_23)[name = string("nv_flat_7")]; + tensor var_2281 = mul(x = var_2075, y = var_1194)[name = string("op_2281")]; + tensor var_2282 = mul(x = nk_flat_7, y = update_mask_1)[name = string("op_2282")]; + tensor key_cache_17 = add(x = var_2281, y = var_2282)[name = string("key_cache_17")]; + tensor var_2288 = mul(x = var_2095, y = var_1194)[name = string("op_2288")]; + tensor var_2289 = mul(x = nv_flat_7, y = update_mask_1)[name = string("op_2289")]; + tensor value_cache_17 = add(x = var_2288, y = var_2289)[name = string("value_cache_17")]; + tensor kc_19_axes_0 = const()[name = string("kc_19_axes_0"), val = tensor([2])]; + tensor kc_19 = squeeze(axes = kc_19_axes_0, x = key_cache_17)[name = string("kc_19")]; + tensor var_2298 = const()[name = string("op_2298"), val = tensor([1, 8, 128, 256])]; + tensor kc_21 = reshape(shape = var_2298, x = kc_19)[name = string("kc_21")]; + tensor vc_19_axes_0 = const()[name = string("vc_19_axes_0"), val = tensor([2])]; + tensor vc_19 = squeeze(axes = vc_19_axes_0, x = value_cache_17)[name = string("vc_19")]; + tensor var_2306 = const()[name = string("op_2306"), val = tensor([1, 8, 128, 256])]; + tensor vc_21 = reshape(shape = var_2306, x = vc_19)[name = string("vc_21")]; + tensor var_2309_axes_0 = const()[name = string("op_2309_axes_0"), val = tensor([2])]; + tensor var_2309 = expand_dims(axes = var_2309_axes_0, x = kc_21)[name = string("op_2309")]; + tensor var_2317_reps_0 = const()[name = string("op_2317_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_2317 = tile(reps = var_2317_reps_0, x = var_2309)[name = string("op_2317")]; + tensor var_2322 = const()[name = string("op_2322"), val = tensor([1, 16, 128, 256])]; + tensor kc_23 = reshape(shape = var_2322, x = var_2317)[name = string("kc_23")]; + tensor var_2325_axes_0 = const()[name = string("op_2325_axes_0"), val = tensor([2])]; + tensor var_2325 = expand_dims(axes = var_2325_axes_0, x = vc_21)[name = string("op_2325")]; + tensor var_2333_reps_0 = const()[name = string("op_2333_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_2333 = tile(reps = var_2333_reps_0, x = var_2325)[name = string("op_2333")]; + tensor var_2338 = const()[name = string("op_2338"), val = tensor([1, 16, 128, 256])]; + tensor vc_23 = reshape(shape = var_2338, x = var_2333)[name = string("vc_23")]; + bool var_2340_transpose_x_0 = const()[name = string("op_2340_transpose_x_0"), val = bool(false)]; + bool var_2340_transpose_y_0 = const()[name = string("op_2340_transpose_y_0"), val = bool(false)]; + tensor var_2340 = matmul(transpose_x = var_2340_transpose_x_0, transpose_y = var_2340_transpose_y_0, x = q_31, y = kc_23)[name = string("op_2340")]; + fp32 _inversed_attn_weights_25_y_0 = const()[name = string("_inversed_attn_weights_25_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_25 = mul(x = var_2340, y = _inversed_attn_weights_25_y_0)[name = string("_inversed_attn_weights_25")]; + tensor attn_weights_27 = add(x = _inversed_attn_weights_25, y = mask_1)[name = string("attn_weights_27")]; + int32 var_2354 = const()[name = string("op_2354"), val = int32(-1)]; + tensor attn_weights_31 = softmax(axis = var_2354, x = attn_weights_27)[name = string("attn_weights_31")]; + bool attn_output_13_transpose_x_1 = const()[name = string("attn_output_13_transpose_x_1"), val = bool(false)]; + bool attn_output_13_transpose_y_1 = const()[name = string("attn_output_13_transpose_y_1"), val = bool(true)]; + tensor attn_output_13 = matmul(transpose_x = attn_output_13_transpose_x_1, transpose_y = attn_output_13_transpose_y_1, x = attn_weights_31, y = vc_23)[name = string("attn_output_13")]; + tensor var_2363_perm_0 = const()[name = string("op_2363_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2367 = const()[name = string("op_2367"), val = tensor([1, 1, -1])]; + tensor var_2363 = transpose(perm = var_2363_perm_0, x = attn_output_13)[name = string("transpose_96")]; + tensor input_33 = reshape(shape = var_2367, x = var_2363)[name = string("input_33")]; + tensor attn_output_15 = linear(bias = linear_1_bias_0, weight = layers_3_self_attn_o_proj_weight, x = input_33)[name = string("linear_24")]; + tensor var_2373_axes_0 = const()[name = string("op_2373_axes_0"), val = tensor([0])]; + tensor var_2373 = squeeze(axes = var_2373_axes_0, x = attn_output_15)[name = string("op_2373")]; + tensor var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor([0])]; + tensor var_2375 = squeeze(axes = var_2375_axes_0, x = var_2373)[name = string("op_2375")]; + tensor var_2377_axes_0 = const()[name = string("op_2377_axes_0"), val = tensor([-1])]; + tensor var_2377 = expand_dims(axes = var_2377_axes_0, x = var_2375)[name = string("op_2377")]; + tensor attn_4d_7_axes_0 = const()[name = string("attn_4d_7_axes_0"), val = tensor([-1])]; + tensor attn_4d_7 = expand_dims(axes = attn_4d_7_axes_0, x = var_2377)[name = string("attn_4d_7")]; + tensor hidden_13 = add(x = hidden_11, y = attn_4d_7)[name = string("hidden_13")]; + tensor var_2383_axes_0 = const()[name = string("op_2383_axes_0"), val = tensor([-1])]; + tensor var_2383 = squeeze(axes = var_2383_axes_0, x = hidden_13)[name = string("op_2383")]; + tensor var_2385_axes_0 = const()[name = string("op_2385_axes_0"), val = tensor([-1])]; + tensor var_2385 = squeeze(axes = var_2385_axes_0, x = var_2383)[name = string("op_2385")]; + tensor hidden_states_91_axes_0 = const()[name = string("hidden_states_91_axes_0"), val = tensor([0])]; + tensor hidden_states_91 = expand_dims(axes = hidden_states_91_axes_0, x = var_2385)[name = string("hidden_states_91")]; + fp32 var_2391_promoted = const()[name = string("op_2391_promoted"), val = fp32(0x1p+1)]; + tensor var_2397 = pow(x = hidden_states_91, y = var_2391_promoted)[name = string("op_2397")]; + tensor variance_31_axes_0 = const()[name = string("variance_31_axes_0"), val = tensor([-1])]; + bool variance_31_keep_dims_0 = const()[name = string("variance_31_keep_dims_0"), val = bool(true)]; + tensor variance_31 = reduce_mean(axes = variance_31_axes_0, keep_dims = variance_31_keep_dims_0, x = var_2397)[name = string("variance_31")]; + fp32 var_2400 = const()[name = string("op_2400"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2401 = add(x = variance_31, y = var_2400)[name = string("op_2401")]; + fp32 var_2402_epsilon_0 = const()[name = string("op_2402_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2402 = rsqrt(epsilon = var_2402_epsilon_0, x = var_2401)[name = string("op_2402")]; + tensor hidden_states_95 = mul(x = hidden_states_91, y = var_2402)[name = string("hidden_states_95")]; + tensor const_40 = const()[name = string("const_40"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774262080)))]; + tensor input_35 = mul(x = const_40, y = hidden_states_95)[name = string("input_35")]; + tensor input_37 = linear(bias = linear_4_bias_0, weight = layers_3_mlp_gate_proj_weight, x = input_35)[name = string("linear_25")]; + tensor var_2412 = silu(x = input_37)[name = string("op_2412")]; + tensor var_2414 = linear(bias = linear_4_bias_0, weight = layers_3_mlp_up_proj_weight, x = input_35)[name = string("linear_26")]; + tensor input_39 = mul(x = var_2412, y = var_2414)[name = string("input_39")]; + tensor mlp_out_7 = linear(bias = linear_1_bias_0, weight = layers_3_mlp_down_proj_weight, x = input_39)[name = string("linear_27")]; + tensor var_2419_axes_0 = const()[name = string("op_2419_axes_0"), val = tensor([0])]; + tensor var_2419 = squeeze(axes = var_2419_axes_0, x = mlp_out_7)[name = string("op_2419")]; + tensor var_2421_axes_0 = const()[name = string("op_2421_axes_0"), val = tensor([0])]; + tensor var_2421 = squeeze(axes = var_2421_axes_0, x = var_2419)[name = string("op_2421")]; + tensor var_2423_axes_0 = const()[name = string("op_2423_axes_0"), val = tensor([-1])]; + tensor var_2423 = expand_dims(axes = var_2423_axes_0, x = var_2421)[name = string("op_2423")]; + tensor mlp_4d_7_axes_0 = const()[name = string("mlp_4d_7_axes_0"), val = tensor([-1])]; + tensor mlp_4d_7 = expand_dims(axes = mlp_4d_7_axes_0, x = var_2423)[name = string("mlp_4d_7")]; + tensor hidden_15 = add(x = hidden_13, y = mlp_4d_7)[name = string("hidden_15")]; + tensor var_2437_begin_0 = const()[name = string("op_2437_begin_0"), val = tensor([0, 4096, 0, 0])]; + tensor var_2437_end_0 = const()[name = string("op_2437_end_0"), val = tensor([1, 5120, 1, 256])]; + tensor var_2437_end_mask_0 = const()[name = string("op_2437_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2437 = slice_by_index(begin = var_2437_begin_0, end = var_2437_end_0, end_mask = var_2437_end_mask_0, x = cast_3)[name = string("op_2437")]; + tensor var_2457_begin_0 = const()[name = string("op_2457_begin_0"), val = tensor([0, 4096, 0, 0])]; + tensor var_2457_end_0 = const()[name = string("op_2457_end_0"), val = tensor([1, 5120, 1, 256])]; + tensor var_2457_end_mask_0 = const()[name = string("op_2457_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2457 = slice_by_index(begin = var_2457_begin_0, end = var_2457_end_0, end_mask = var_2457_end_mask_0, x = cast_4)[name = string("op_2457")]; + tensor var_2469_axes_0 = const()[name = string("op_2469_axes_0"), val = tensor([-1])]; + tensor var_2469 = squeeze(axes = var_2469_axes_0, x = hidden_15)[name = string("op_2469")]; + tensor var_2471_axes_0 = const()[name = string("op_2471_axes_0"), val = tensor([-1])]; + tensor var_2471 = squeeze(axes = var_2471_axes_0, x = var_2469)[name = string("op_2471")]; + tensor hidden_states_97_axes_0 = const()[name = string("hidden_states_97_axes_0"), val = tensor([0])]; + tensor hidden_states_97 = expand_dims(axes = hidden_states_97_axes_0, x = var_2471)[name = string("hidden_states_97")]; + fp32 var_2477_promoted = const()[name = string("op_2477_promoted"), val = fp32(0x1p+1)]; + tensor var_2483 = pow(x = hidden_states_97, y = var_2477_promoted)[name = string("op_2483")]; + tensor variance_33_axes_0 = const()[name = string("variance_33_axes_0"), val = tensor([-1])]; + bool variance_33_keep_dims_0 = const()[name = string("variance_33_keep_dims_0"), val = bool(true)]; + tensor variance_33 = reduce_mean(axes = variance_33_axes_0, keep_dims = variance_33_keep_dims_0, x = var_2483)[name = string("variance_33")]; + fp32 var_2486 = const()[name = string("op_2486"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2487 = add(x = variance_33, y = var_2486)[name = string("op_2487")]; + fp32 var_2488_epsilon_0 = const()[name = string("op_2488_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2488 = rsqrt(epsilon = var_2488_epsilon_0, x = var_2487)[name = string("op_2488")]; + tensor hidden_states_101 = mul(x = hidden_states_97, y = var_2488)[name = string("hidden_states_101")]; + tensor const_41 = const()[name = string("const_41"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774266240)))]; + tensor input_41 = mul(x = const_41, y = hidden_states_101)[name = string("input_41")]; + tensor q_33 = linear(bias = linear_0_bias_0, weight = layers_4_self_attn_q_proj_weight, x = input_41)[name = string("linear_28")]; + tensor k_33 = linear(bias = linear_1_bias_0, weight = layers_4_self_attn_k_proj_weight, x = input_41)[name = string("linear_29")]; + tensor v_25 = linear(bias = linear_1_bias_0, weight = layers_4_self_attn_v_proj_weight, x = input_41)[name = string("linear_30")]; + tensor var_2505 = const()[name = string("op_2505"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_103 = reshape(shape = var_2505, x = q_33)[name = string("hidden_states_103")]; + tensor var_2511 = const()[name = string("op_2511"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_109 = reshape(shape = var_2511, x = k_33)[name = string("hidden_states_109")]; + tensor var_2517 = const()[name = string("op_2517"), val = tensor([1, 1, 8, 128])]; + tensor v_27 = reshape(shape = var_2517, x = v_25)[name = string("v_27")]; + fp32 var_2522_promoted = const()[name = string("op_2522_promoted"), val = fp32(0x1p+1)]; + tensor var_2528 = pow(x = hidden_states_103, y = var_2522_promoted)[name = string("op_2528")]; + tensor variance_35_axes_0 = const()[name = string("variance_35_axes_0"), val = tensor([-1])]; + bool variance_35_keep_dims_0 = const()[name = string("variance_35_keep_dims_0"), val = bool(true)]; + tensor variance_35 = reduce_mean(axes = variance_35_axes_0, keep_dims = variance_35_keep_dims_0, x = var_2528)[name = string("variance_35")]; + fp32 var_2531 = const()[name = string("op_2531"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2532 = add(x = variance_35, y = var_2531)[name = string("op_2532")]; + fp32 var_2533_epsilon_0 = const()[name = string("op_2533_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2533 = rsqrt(epsilon = var_2533_epsilon_0, x = var_2532)[name = string("op_2533")]; + tensor hidden_states_107 = mul(x = hidden_states_103, y = var_2533)[name = string("hidden_states_107")]; + tensor const_42 = const()[name = string("const_42"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774270400)))]; + tensor q_35 = mul(x = const_42, y = hidden_states_107)[name = string("q_35")]; + fp32 var_2540_promoted = const()[name = string("op_2540_promoted"), val = fp32(0x1p+1)]; + tensor var_2546 = pow(x = hidden_states_109, y = var_2540_promoted)[name = string("op_2546")]; + tensor variance_37_axes_0 = const()[name = string("variance_37_axes_0"), val = tensor([-1])]; + bool variance_37_keep_dims_0 = const()[name = string("variance_37_keep_dims_0"), val = bool(true)]; + tensor variance_37 = reduce_mean(axes = variance_37_axes_0, keep_dims = variance_37_keep_dims_0, x = var_2546)[name = string("variance_37")]; + fp32 var_2549 = const()[name = string("op_2549"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2550 = add(x = variance_37, y = var_2549)[name = string("op_2550")]; + fp32 var_2551_epsilon_0 = const()[name = string("op_2551_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2551 = rsqrt(epsilon = var_2551_epsilon_0, x = var_2550)[name = string("op_2551")]; + tensor hidden_states_113 = mul(x = hidden_states_109, y = var_2551)[name = string("hidden_states_113")]; + tensor const_43 = const()[name = string("const_43"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774270976)))]; + tensor k_35 = mul(x = const_43, y = hidden_states_113)[name = string("k_35")]; + tensor q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_37 = transpose(perm = q_37_perm_0, x = q_35)[name = string("transpose_95")]; + tensor var_2568 = mul(x = q_37, y = cos_3)[name = string("op_2568")]; + tensor x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_17 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_37)[name = string("x1_17")]; + tensor x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_17 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_37)[name = string("x2_17")]; + fp32 const_46_promoted = const()[name = string("const_46_promoted"), val = fp32(-0x1p+0)]; + tensor var_2589 = mul(x = x2_17, y = const_46_promoted)[name = string("op_2589")]; + int32 var_2591 = const()[name = string("op_2591"), val = int32(-1)]; + bool var_2592_interleave_0 = const()[name = string("op_2592_interleave_0"), val = bool(false)]; + tensor var_2592 = concat(axis = var_2591, interleave = var_2592_interleave_0, values = (var_2589, x1_17))[name = string("op_2592")]; + tensor var_2593 = mul(x = var_2592, y = sin_3)[name = string("op_2593")]; + tensor q_39 = add(x = var_2568, y = var_2593)[name = string("q_39")]; + tensor k_37 = transpose(perm = k_37_perm_0, x = k_35)[name = string("transpose_94")]; + tensor var_2596 = mul(x = k_37, y = cos_3)[name = string("op_2596")]; + tensor x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_19 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_37)[name = string("x1_19")]; + tensor x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_19 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_37)[name = string("x2_19")]; + fp32 const_49_promoted = const()[name = string("const_49_promoted"), val = fp32(-0x1p+0)]; + tensor var_2617 = mul(x = x2_19, y = const_49_promoted)[name = string("op_2617")]; + int32 var_2619 = const()[name = string("op_2619"), val = int32(-1)]; + bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)]; + tensor var_2620 = concat(axis = var_2619, interleave = var_2620_interleave_0, values = (var_2617, x1_19))[name = string("op_2620")]; + tensor var_2621 = mul(x = var_2620, y = sin_3)[name = string("op_2621")]; + tensor k_39 = add(x = var_2596, y = var_2621)[name = string("k_39")]; + tensor var_2628 = const()[name = string("op_2628"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_9 = reshape(shape = var_2628, x = k_39)[name = string("nk_flat_9")]; + tensor var_2634 = const()[name = string("op_2634"), val = tensor([1, 1024, 1, 1])]; + tensor v_29 = transpose(perm = v_29_perm_0, x = v_27)[name = string("transpose_93")]; + tensor nv_flat_9 = reshape(shape = var_2634, x = v_29)[name = string("nv_flat_9")]; + tensor var_2643 = mul(x = var_2437, y = var_1194)[name = string("op_2643")]; + tensor var_2644 = mul(x = nk_flat_9, y = update_mask_1)[name = string("op_2644")]; + tensor key_cache_21 = add(x = var_2643, y = var_2644)[name = string("key_cache_21")]; + tensor var_2650 = mul(x = var_2457, y = var_1194)[name = string("op_2650")]; + tensor var_2651 = mul(x = nv_flat_9, y = update_mask_1)[name = string("op_2651")]; + tensor value_cache_21 = add(x = var_2650, y = var_2651)[name = string("value_cache_21")]; + tensor kc_25_axes_0 = const()[name = string("kc_25_axes_0"), val = tensor([2])]; + tensor kc_25 = squeeze(axes = kc_25_axes_0, x = key_cache_21)[name = string("kc_25")]; + tensor var_2660 = const()[name = string("op_2660"), val = tensor([1, 8, 128, 256])]; + tensor kc_27 = reshape(shape = var_2660, x = kc_25)[name = string("kc_27")]; + tensor vc_25_axes_0 = const()[name = string("vc_25_axes_0"), val = tensor([2])]; + tensor vc_25 = squeeze(axes = vc_25_axes_0, x = value_cache_21)[name = string("vc_25")]; + tensor var_2668 = const()[name = string("op_2668"), val = tensor([1, 8, 128, 256])]; + tensor vc_27 = reshape(shape = var_2668, x = vc_25)[name = string("vc_27")]; + tensor var_2671_axes_0 = const()[name = string("op_2671_axes_0"), val = tensor([2])]; + tensor var_2671 = expand_dims(axes = var_2671_axes_0, x = kc_27)[name = string("op_2671")]; + tensor var_2679_reps_0 = const()[name = string("op_2679_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_2679 = tile(reps = var_2679_reps_0, x = var_2671)[name = string("op_2679")]; + tensor var_2684 = const()[name = string("op_2684"), val = tensor([1, 16, 128, 256])]; + tensor kc_29 = reshape(shape = var_2684, x = var_2679)[name = string("kc_29")]; + tensor var_2687_axes_0 = const()[name = string("op_2687_axes_0"), val = tensor([2])]; + tensor var_2687 = expand_dims(axes = var_2687_axes_0, x = vc_27)[name = string("op_2687")]; + tensor var_2695_reps_0 = const()[name = string("op_2695_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_2695 = tile(reps = var_2695_reps_0, x = var_2687)[name = string("op_2695")]; + tensor var_2700 = const()[name = string("op_2700"), val = tensor([1, 16, 128, 256])]; + tensor vc_29 = reshape(shape = var_2700, x = var_2695)[name = string("vc_29")]; + bool var_2702_transpose_x_0 = const()[name = string("op_2702_transpose_x_0"), val = bool(false)]; + bool var_2702_transpose_y_0 = const()[name = string("op_2702_transpose_y_0"), val = bool(false)]; + tensor var_2702 = matmul(transpose_x = var_2702_transpose_x_0, transpose_y = var_2702_transpose_y_0, x = q_39, y = kc_29)[name = string("op_2702")]; + fp32 _inversed_attn_weights_33_y_0 = const()[name = string("_inversed_attn_weights_33_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_33 = mul(x = var_2702, y = _inversed_attn_weights_33_y_0)[name = string("_inversed_attn_weights_33")]; + tensor attn_weights_35 = add(x = _inversed_attn_weights_33, y = mask_1)[name = string("attn_weights_35")]; + int32 var_2716 = const()[name = string("op_2716"), val = int32(-1)]; + tensor attn_weights_39 = softmax(axis = var_2716, x = attn_weights_35)[name = string("attn_weights_39")]; + bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(false)]; + bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(true)]; + tensor attn_output_17 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_39, y = vc_29)[name = string("attn_output_17")]; + tensor var_2725_perm_0 = const()[name = string("op_2725_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2729 = const()[name = string("op_2729"), val = tensor([1, 1, -1])]; + tensor var_2725 = transpose(perm = var_2725_perm_0, x = attn_output_17)[name = string("transpose_92")]; + tensor input_43 = reshape(shape = var_2729, x = var_2725)[name = string("input_43")]; + tensor attn_output_19 = linear(bias = linear_1_bias_0, weight = layers_4_self_attn_o_proj_weight, x = input_43)[name = string("linear_31")]; + tensor var_2735_axes_0 = const()[name = string("op_2735_axes_0"), val = tensor([0])]; + tensor var_2735 = squeeze(axes = var_2735_axes_0, x = attn_output_19)[name = string("op_2735")]; + tensor var_2737_axes_0 = const()[name = string("op_2737_axes_0"), val = tensor([0])]; + tensor var_2737 = squeeze(axes = var_2737_axes_0, x = var_2735)[name = string("op_2737")]; + tensor var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor([-1])]; + tensor var_2739 = expand_dims(axes = var_2739_axes_0, x = var_2737)[name = string("op_2739")]; + tensor attn_4d_9_axes_0 = const()[name = string("attn_4d_9_axes_0"), val = tensor([-1])]; + tensor attn_4d_9 = expand_dims(axes = attn_4d_9_axes_0, x = var_2739)[name = string("attn_4d_9")]; + tensor hidden_17 = add(x = hidden_15, y = attn_4d_9)[name = string("hidden_17")]; + tensor var_2745_axes_0 = const()[name = string("op_2745_axes_0"), val = tensor([-1])]; + tensor var_2745 = squeeze(axes = var_2745_axes_0, x = hidden_17)[name = string("op_2745")]; + tensor var_2747_axes_0 = const()[name = string("op_2747_axes_0"), val = tensor([-1])]; + tensor var_2747 = squeeze(axes = var_2747_axes_0, x = var_2745)[name = string("op_2747")]; + tensor hidden_states_115_axes_0 = const()[name = string("hidden_states_115_axes_0"), val = tensor([0])]; + tensor hidden_states_115 = expand_dims(axes = hidden_states_115_axes_0, x = var_2747)[name = string("hidden_states_115")]; + fp32 var_2753_promoted = const()[name = string("op_2753_promoted"), val = fp32(0x1p+1)]; + tensor var_2759 = pow(x = hidden_states_115, y = var_2753_promoted)[name = string("op_2759")]; + tensor variance_39_axes_0 = const()[name = string("variance_39_axes_0"), val = tensor([-1])]; + bool variance_39_keep_dims_0 = const()[name = string("variance_39_keep_dims_0"), val = bool(true)]; + tensor variance_39 = reduce_mean(axes = variance_39_axes_0, keep_dims = variance_39_keep_dims_0, x = var_2759)[name = string("variance_39")]; + fp32 var_2762 = const()[name = string("op_2762"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2763 = add(x = variance_39, y = var_2762)[name = string("op_2763")]; + fp32 var_2764_epsilon_0 = const()[name = string("op_2764_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2764 = rsqrt(epsilon = var_2764_epsilon_0, x = var_2763)[name = string("op_2764")]; + tensor hidden_states_119 = mul(x = hidden_states_115, y = var_2764)[name = string("hidden_states_119")]; + tensor const_50 = const()[name = string("const_50"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774271552)))]; + tensor input_45 = mul(x = const_50, y = hidden_states_119)[name = string("input_45")]; + tensor input_47 = linear(bias = linear_4_bias_0, weight = layers_4_mlp_gate_proj_weight, x = input_45)[name = string("linear_32")]; + tensor var_2774 = silu(x = input_47)[name = string("op_2774")]; + tensor var_2776 = linear(bias = linear_4_bias_0, weight = layers_4_mlp_up_proj_weight, x = input_45)[name = string("linear_33")]; + tensor input_49 = mul(x = var_2774, y = var_2776)[name = string("input_49")]; + tensor mlp_out_9 = linear(bias = linear_1_bias_0, weight = layers_4_mlp_down_proj_weight, x = input_49)[name = string("linear_34")]; + tensor var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor([0])]; + tensor var_2781 = squeeze(axes = var_2781_axes_0, x = mlp_out_9)[name = string("op_2781")]; + tensor var_2783_axes_0 = const()[name = string("op_2783_axes_0"), val = tensor([0])]; + tensor var_2783 = squeeze(axes = var_2783_axes_0, x = var_2781)[name = string("op_2783")]; + tensor var_2785_axes_0 = const()[name = string("op_2785_axes_0"), val = tensor([-1])]; + tensor var_2785 = expand_dims(axes = var_2785_axes_0, x = var_2783)[name = string("op_2785")]; + tensor mlp_4d_9_axes_0 = const()[name = string("mlp_4d_9_axes_0"), val = tensor([-1])]; + tensor mlp_4d_9 = expand_dims(axes = mlp_4d_9_axes_0, x = var_2785)[name = string("mlp_4d_9")]; + tensor hidden_19 = add(x = hidden_17, y = mlp_4d_9)[name = string("hidden_19")]; + tensor var_2799_begin_0 = const()[name = string("op_2799_begin_0"), val = tensor([0, 5120, 0, 0])]; + tensor var_2799_end_0 = const()[name = string("op_2799_end_0"), val = tensor([1, 6144, 1, 256])]; + tensor var_2799_end_mask_0 = const()[name = string("op_2799_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2799 = slice_by_index(begin = var_2799_begin_0, end = var_2799_end_0, end_mask = var_2799_end_mask_0, x = cast_3)[name = string("op_2799")]; + tensor var_2819_begin_0 = const()[name = string("op_2819_begin_0"), val = tensor([0, 5120, 0, 0])]; + tensor var_2819_end_0 = const()[name = string("op_2819_end_0"), val = tensor([1, 6144, 1, 256])]; + tensor var_2819_end_mask_0 = const()[name = string("op_2819_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2819 = slice_by_index(begin = var_2819_begin_0, end = var_2819_end_0, end_mask = var_2819_end_mask_0, x = cast_4)[name = string("op_2819")]; + tensor var_2831_axes_0 = const()[name = string("op_2831_axes_0"), val = tensor([-1])]; + tensor var_2831 = squeeze(axes = var_2831_axes_0, x = hidden_19)[name = string("op_2831")]; + tensor var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor([-1])]; + tensor var_2833 = squeeze(axes = var_2833_axes_0, x = var_2831)[name = string("op_2833")]; + tensor hidden_states_121_axes_0 = const()[name = string("hidden_states_121_axes_0"), val = tensor([0])]; + tensor hidden_states_121 = expand_dims(axes = hidden_states_121_axes_0, x = var_2833)[name = string("hidden_states_121")]; + fp32 var_2839_promoted = const()[name = string("op_2839_promoted"), val = fp32(0x1p+1)]; + tensor var_2845 = pow(x = hidden_states_121, y = var_2839_promoted)[name = string("op_2845")]; + tensor variance_41_axes_0 = const()[name = string("variance_41_axes_0"), val = tensor([-1])]; + bool variance_41_keep_dims_0 = const()[name = string("variance_41_keep_dims_0"), val = bool(true)]; + tensor variance_41 = reduce_mean(axes = variance_41_axes_0, keep_dims = variance_41_keep_dims_0, x = var_2845)[name = string("variance_41")]; + fp32 var_2848 = const()[name = string("op_2848"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2849 = add(x = variance_41, y = var_2848)[name = string("op_2849")]; + fp32 var_2850_epsilon_0 = const()[name = string("op_2850_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2850 = rsqrt(epsilon = var_2850_epsilon_0, x = var_2849)[name = string("op_2850")]; + tensor hidden_states_125 = mul(x = hidden_states_121, y = var_2850)[name = string("hidden_states_125")]; + tensor const_51 = const()[name = string("const_51"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774275712)))]; + tensor input_51 = mul(x = const_51, y = hidden_states_125)[name = string("input_51")]; + tensor q_41 = linear(bias = linear_0_bias_0, weight = layers_5_self_attn_q_proj_weight, x = input_51)[name = string("linear_35")]; + tensor k_41 = linear(bias = linear_1_bias_0, weight = layers_5_self_attn_k_proj_weight, x = input_51)[name = string("linear_36")]; + tensor v_31 = linear(bias = linear_1_bias_0, weight = layers_5_self_attn_v_proj_weight, x = input_51)[name = string("linear_37")]; + tensor var_2867 = const()[name = string("op_2867"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_127 = reshape(shape = var_2867, x = q_41)[name = string("hidden_states_127")]; + tensor var_2873 = const()[name = string("op_2873"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_133 = reshape(shape = var_2873, x = k_41)[name = string("hidden_states_133")]; + tensor var_2879 = const()[name = string("op_2879"), val = tensor([1, 1, 8, 128])]; + tensor v_33 = reshape(shape = var_2879, x = v_31)[name = string("v_33")]; + fp32 var_2884_promoted = const()[name = string("op_2884_promoted"), val = fp32(0x1p+1)]; + tensor var_2890 = pow(x = hidden_states_127, y = var_2884_promoted)[name = string("op_2890")]; + tensor variance_43_axes_0 = const()[name = string("variance_43_axes_0"), val = tensor([-1])]; + bool variance_43_keep_dims_0 = const()[name = string("variance_43_keep_dims_0"), val = bool(true)]; + tensor variance_43 = reduce_mean(axes = variance_43_axes_0, keep_dims = variance_43_keep_dims_0, x = var_2890)[name = string("variance_43")]; + fp32 var_2893 = const()[name = string("op_2893"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2894 = add(x = variance_43, y = var_2893)[name = string("op_2894")]; + fp32 var_2895_epsilon_0 = const()[name = string("op_2895_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2895 = rsqrt(epsilon = var_2895_epsilon_0, x = var_2894)[name = string("op_2895")]; + tensor hidden_states_131 = mul(x = hidden_states_127, y = var_2895)[name = string("hidden_states_131")]; + tensor const_52 = const()[name = string("const_52"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774279872)))]; + tensor q_43 = mul(x = const_52, y = hidden_states_131)[name = string("q_43")]; + fp32 var_2902_promoted = const()[name = string("op_2902_promoted"), val = fp32(0x1p+1)]; + tensor var_2908 = pow(x = hidden_states_133, y = var_2902_promoted)[name = string("op_2908")]; + tensor variance_45_axes_0 = const()[name = string("variance_45_axes_0"), val = tensor([-1])]; + bool variance_45_keep_dims_0 = const()[name = string("variance_45_keep_dims_0"), val = bool(true)]; + tensor variance_45 = reduce_mean(axes = variance_45_axes_0, keep_dims = variance_45_keep_dims_0, x = var_2908)[name = string("variance_45")]; + fp32 var_2911 = const()[name = string("op_2911"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_2912 = add(x = variance_45, y = var_2911)[name = string("op_2912")]; + fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_2913 = rsqrt(epsilon = var_2913_epsilon_0, x = var_2912)[name = string("op_2913")]; + tensor hidden_states_137 = mul(x = hidden_states_133, y = var_2913)[name = string("hidden_states_137")]; + tensor const_53 = const()[name = string("const_53"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774280448)))]; + tensor k_43 = mul(x = const_53, y = hidden_states_137)[name = string("k_43")]; + tensor q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_45 = transpose(perm = q_45_perm_0, x = q_43)[name = string("transpose_91")]; + tensor var_2930 = mul(x = q_45, y = cos_3)[name = string("op_2930")]; + tensor x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_21 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_45)[name = string("x1_21")]; + tensor x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_21 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_45)[name = string("x2_21")]; + fp32 const_56_promoted = const()[name = string("const_56_promoted"), val = fp32(-0x1p+0)]; + tensor var_2951 = mul(x = x2_21, y = const_56_promoted)[name = string("op_2951")]; + int32 var_2953 = const()[name = string("op_2953"), val = int32(-1)]; + bool var_2954_interleave_0 = const()[name = string("op_2954_interleave_0"), val = bool(false)]; + tensor var_2954 = concat(axis = var_2953, interleave = var_2954_interleave_0, values = (var_2951, x1_21))[name = string("op_2954")]; + tensor var_2955 = mul(x = var_2954, y = sin_3)[name = string("op_2955")]; + tensor q_47 = add(x = var_2930, y = var_2955)[name = string("q_47")]; + tensor k_45 = transpose(perm = k_45_perm_0, x = k_43)[name = string("transpose_90")]; + tensor var_2958 = mul(x = k_45, y = cos_3)[name = string("op_2958")]; + tensor x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_23 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_45)[name = string("x1_23")]; + tensor x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_23 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_45)[name = string("x2_23")]; + fp32 const_59_promoted = const()[name = string("const_59_promoted"), val = fp32(-0x1p+0)]; + tensor var_2979 = mul(x = x2_23, y = const_59_promoted)[name = string("op_2979")]; + int32 var_2981 = const()[name = string("op_2981"), val = int32(-1)]; + bool var_2982_interleave_0 = const()[name = string("op_2982_interleave_0"), val = bool(false)]; + tensor var_2982 = concat(axis = var_2981, interleave = var_2982_interleave_0, values = (var_2979, x1_23))[name = string("op_2982")]; + tensor var_2983 = mul(x = var_2982, y = sin_3)[name = string("op_2983")]; + tensor k_47 = add(x = var_2958, y = var_2983)[name = string("k_47")]; + tensor var_2990 = const()[name = string("op_2990"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_11 = reshape(shape = var_2990, x = k_47)[name = string("nk_flat_11")]; + tensor var_2996 = const()[name = string("op_2996"), val = tensor([1, 1024, 1, 1])]; + tensor v_35 = transpose(perm = v_35_perm_0, x = v_33)[name = string("transpose_89")]; + tensor nv_flat_11 = reshape(shape = var_2996, x = v_35)[name = string("nv_flat_11")]; + tensor var_3005 = mul(x = var_2799, y = var_1194)[name = string("op_3005")]; + tensor var_3006 = mul(x = nk_flat_11, y = update_mask_1)[name = string("op_3006")]; + tensor key_cache_25 = add(x = var_3005, y = var_3006)[name = string("key_cache_25")]; + tensor var_3012 = mul(x = var_2819, y = var_1194)[name = string("op_3012")]; + tensor var_3013 = mul(x = nv_flat_11, y = update_mask_1)[name = string("op_3013")]; + tensor value_cache_25 = add(x = var_3012, y = var_3013)[name = string("value_cache_25")]; + tensor kc_31_axes_0 = const()[name = string("kc_31_axes_0"), val = tensor([2])]; + tensor kc_31 = squeeze(axes = kc_31_axes_0, x = key_cache_25)[name = string("kc_31")]; + tensor var_3022 = const()[name = string("op_3022"), val = tensor([1, 8, 128, 256])]; + tensor kc_33 = reshape(shape = var_3022, x = kc_31)[name = string("kc_33")]; + tensor vc_31_axes_0 = const()[name = string("vc_31_axes_0"), val = tensor([2])]; + tensor vc_31 = squeeze(axes = vc_31_axes_0, x = value_cache_25)[name = string("vc_31")]; + tensor var_3030 = const()[name = string("op_3030"), val = tensor([1, 8, 128, 256])]; + tensor vc_33 = reshape(shape = var_3030, x = vc_31)[name = string("vc_33")]; + tensor var_3033_axes_0 = const()[name = string("op_3033_axes_0"), val = tensor([2])]; + tensor var_3033 = expand_dims(axes = var_3033_axes_0, x = kc_33)[name = string("op_3033")]; + tensor var_3041_reps_0 = const()[name = string("op_3041_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_3041 = tile(reps = var_3041_reps_0, x = var_3033)[name = string("op_3041")]; + tensor var_3046 = const()[name = string("op_3046"), val = tensor([1, 16, 128, 256])]; + tensor kc_35 = reshape(shape = var_3046, x = var_3041)[name = string("kc_35")]; + tensor var_3049_axes_0 = const()[name = string("op_3049_axes_0"), val = tensor([2])]; + tensor var_3049 = expand_dims(axes = var_3049_axes_0, x = vc_33)[name = string("op_3049")]; + tensor var_3057_reps_0 = const()[name = string("op_3057_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_3057 = tile(reps = var_3057_reps_0, x = var_3049)[name = string("op_3057")]; + tensor var_3062 = const()[name = string("op_3062"), val = tensor([1, 16, 128, 256])]; + tensor vc_35 = reshape(shape = var_3062, x = var_3057)[name = string("vc_35")]; + bool var_3064_transpose_x_0 = const()[name = string("op_3064_transpose_x_0"), val = bool(false)]; + bool var_3064_transpose_y_0 = const()[name = string("op_3064_transpose_y_0"), val = bool(false)]; + tensor var_3064 = matmul(transpose_x = var_3064_transpose_x_0, transpose_y = var_3064_transpose_y_0, x = q_47, y = kc_35)[name = string("op_3064")]; + fp32 _inversed_attn_weights_41_y_0 = const()[name = string("_inversed_attn_weights_41_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_41 = mul(x = var_3064, y = _inversed_attn_weights_41_y_0)[name = string("_inversed_attn_weights_41")]; + tensor attn_weights_43 = add(x = _inversed_attn_weights_41, y = mask_1)[name = string("attn_weights_43")]; + int32 var_3078 = const()[name = string("op_3078"), val = int32(-1)]; + tensor attn_weights_47 = softmax(axis = var_3078, x = attn_weights_43)[name = string("attn_weights_47")]; + bool attn_output_21_transpose_x_1 = const()[name = string("attn_output_21_transpose_x_1"), val = bool(false)]; + bool attn_output_21_transpose_y_1 = const()[name = string("attn_output_21_transpose_y_1"), val = bool(true)]; + tensor attn_output_21 = matmul(transpose_x = attn_output_21_transpose_x_1, transpose_y = attn_output_21_transpose_y_1, x = attn_weights_47, y = vc_35)[name = string("attn_output_21")]; + tensor var_3087_perm_0 = const()[name = string("op_3087_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3091 = const()[name = string("op_3091"), val = tensor([1, 1, -1])]; + tensor var_3087 = transpose(perm = var_3087_perm_0, x = attn_output_21)[name = string("transpose_88")]; + tensor input_53 = reshape(shape = var_3091, x = var_3087)[name = string("input_53")]; + tensor attn_output_23 = linear(bias = linear_1_bias_0, weight = layers_5_self_attn_o_proj_weight, x = input_53)[name = string("linear_38")]; + tensor var_3097_axes_0 = const()[name = string("op_3097_axes_0"), val = tensor([0])]; + tensor var_3097 = squeeze(axes = var_3097_axes_0, x = attn_output_23)[name = string("op_3097")]; + tensor var_3099_axes_0 = const()[name = string("op_3099_axes_0"), val = tensor([0])]; + tensor var_3099 = squeeze(axes = var_3099_axes_0, x = var_3097)[name = string("op_3099")]; + tensor var_3101_axes_0 = const()[name = string("op_3101_axes_0"), val = tensor([-1])]; + tensor var_3101 = expand_dims(axes = var_3101_axes_0, x = var_3099)[name = string("op_3101")]; + tensor attn_4d_11_axes_0 = const()[name = string("attn_4d_11_axes_0"), val = tensor([-1])]; + tensor attn_4d_11 = expand_dims(axes = attn_4d_11_axes_0, x = var_3101)[name = string("attn_4d_11")]; + tensor hidden_21 = add(x = hidden_19, y = attn_4d_11)[name = string("hidden_21")]; + tensor var_3107_axes_0 = const()[name = string("op_3107_axes_0"), val = tensor([-1])]; + tensor var_3107 = squeeze(axes = var_3107_axes_0, x = hidden_21)[name = string("op_3107")]; + tensor var_3109_axes_0 = const()[name = string("op_3109_axes_0"), val = tensor([-1])]; + tensor var_3109 = squeeze(axes = var_3109_axes_0, x = var_3107)[name = string("op_3109")]; + tensor hidden_states_139_axes_0 = const()[name = string("hidden_states_139_axes_0"), val = tensor([0])]; + tensor hidden_states_139 = expand_dims(axes = hidden_states_139_axes_0, x = var_3109)[name = string("hidden_states_139")]; + fp32 var_3115_promoted = const()[name = string("op_3115_promoted"), val = fp32(0x1p+1)]; + tensor var_3121 = pow(x = hidden_states_139, y = var_3115_promoted)[name = string("op_3121")]; + tensor variance_47_axes_0 = const()[name = string("variance_47_axes_0"), val = tensor([-1])]; + bool variance_47_keep_dims_0 = const()[name = string("variance_47_keep_dims_0"), val = bool(true)]; + tensor variance_47 = reduce_mean(axes = variance_47_axes_0, keep_dims = variance_47_keep_dims_0, x = var_3121)[name = string("variance_47")]; + fp32 var_3124 = const()[name = string("op_3124"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3125 = add(x = variance_47, y = var_3124)[name = string("op_3125")]; + fp32 var_3126_epsilon_0 = const()[name = string("op_3126_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3126 = rsqrt(epsilon = var_3126_epsilon_0, x = var_3125)[name = string("op_3126")]; + tensor hidden_states_143 = mul(x = hidden_states_139, y = var_3126)[name = string("hidden_states_143")]; + tensor const_60 = const()[name = string("const_60"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774281024)))]; + tensor input_55 = mul(x = const_60, y = hidden_states_143)[name = string("input_55")]; + tensor input_57 = linear(bias = linear_4_bias_0, weight = layers_5_mlp_gate_proj_weight, x = input_55)[name = string("linear_39")]; + tensor var_3136 = silu(x = input_57)[name = string("op_3136")]; + tensor var_3138 = linear(bias = linear_4_bias_0, weight = layers_5_mlp_up_proj_weight, x = input_55)[name = string("linear_40")]; + tensor input_59 = mul(x = var_3136, y = var_3138)[name = string("input_59")]; + tensor mlp_out_11 = linear(bias = linear_1_bias_0, weight = layers_5_mlp_down_proj_weight, x = input_59)[name = string("linear_41")]; + tensor var_3143_axes_0 = const()[name = string("op_3143_axes_0"), val = tensor([0])]; + tensor var_3143 = squeeze(axes = var_3143_axes_0, x = mlp_out_11)[name = string("op_3143")]; + tensor var_3145_axes_0 = const()[name = string("op_3145_axes_0"), val = tensor([0])]; + tensor var_3145 = squeeze(axes = var_3145_axes_0, x = var_3143)[name = string("op_3145")]; + tensor var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor([-1])]; + tensor var_3147 = expand_dims(axes = var_3147_axes_0, x = var_3145)[name = string("op_3147")]; + tensor mlp_4d_11_axes_0 = const()[name = string("mlp_4d_11_axes_0"), val = tensor([-1])]; + tensor mlp_4d_11 = expand_dims(axes = mlp_4d_11_axes_0, x = var_3147)[name = string("mlp_4d_11")]; + tensor hidden_23 = add(x = hidden_21, y = mlp_4d_11)[name = string("hidden_23")]; + tensor var_3161_begin_0 = const()[name = string("op_3161_begin_0"), val = tensor([0, 6144, 0, 0])]; + tensor var_3161_end_0 = const()[name = string("op_3161_end_0"), val = tensor([1, 7168, 1, 256])]; + tensor var_3161_end_mask_0 = const()[name = string("op_3161_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3161 = slice_by_index(begin = var_3161_begin_0, end = var_3161_end_0, end_mask = var_3161_end_mask_0, x = cast_3)[name = string("op_3161")]; + tensor var_3181_begin_0 = const()[name = string("op_3181_begin_0"), val = tensor([0, 6144, 0, 0])]; + tensor var_3181_end_0 = const()[name = string("op_3181_end_0"), val = tensor([1, 7168, 1, 256])]; + tensor var_3181_end_mask_0 = const()[name = string("op_3181_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3181 = slice_by_index(begin = var_3181_begin_0, end = var_3181_end_0, end_mask = var_3181_end_mask_0, x = cast_4)[name = string("op_3181")]; + tensor var_3193_axes_0 = const()[name = string("op_3193_axes_0"), val = tensor([-1])]; + tensor var_3193 = squeeze(axes = var_3193_axes_0, x = hidden_23)[name = string("op_3193")]; + tensor var_3195_axes_0 = const()[name = string("op_3195_axes_0"), val = tensor([-1])]; + tensor var_3195 = squeeze(axes = var_3195_axes_0, x = var_3193)[name = string("op_3195")]; + tensor hidden_states_145_axes_0 = const()[name = string("hidden_states_145_axes_0"), val = tensor([0])]; + tensor hidden_states_145 = expand_dims(axes = hidden_states_145_axes_0, x = var_3195)[name = string("hidden_states_145")]; + fp32 var_3201_promoted = const()[name = string("op_3201_promoted"), val = fp32(0x1p+1)]; + tensor var_3207 = pow(x = hidden_states_145, y = var_3201_promoted)[name = string("op_3207")]; + tensor variance_49_axes_0 = const()[name = string("variance_49_axes_0"), val = tensor([-1])]; + bool variance_49_keep_dims_0 = const()[name = string("variance_49_keep_dims_0"), val = bool(true)]; + tensor variance_49 = reduce_mean(axes = variance_49_axes_0, keep_dims = variance_49_keep_dims_0, x = var_3207)[name = string("variance_49")]; + fp32 var_3210 = const()[name = string("op_3210"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3211 = add(x = variance_49, y = var_3210)[name = string("op_3211")]; + fp32 var_3212_epsilon_0 = const()[name = string("op_3212_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3212 = rsqrt(epsilon = var_3212_epsilon_0, x = var_3211)[name = string("op_3212")]; + tensor hidden_states_149 = mul(x = hidden_states_145, y = var_3212)[name = string("hidden_states_149")]; + tensor const_61 = const()[name = string("const_61"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774285184)))]; + tensor input_61 = mul(x = const_61, y = hidden_states_149)[name = string("input_61")]; + tensor q_49 = linear(bias = linear_0_bias_0, weight = layers_6_self_attn_q_proj_weight, x = input_61)[name = string("linear_42")]; + tensor k_49 = linear(bias = linear_1_bias_0, weight = layers_6_self_attn_k_proj_weight, x = input_61)[name = string("linear_43")]; + tensor v_37 = linear(bias = linear_1_bias_0, weight = layers_6_self_attn_v_proj_weight, x = input_61)[name = string("linear_44")]; + tensor var_3229 = const()[name = string("op_3229"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_151 = reshape(shape = var_3229, x = q_49)[name = string("hidden_states_151")]; + tensor var_3235 = const()[name = string("op_3235"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_157 = reshape(shape = var_3235, x = k_49)[name = string("hidden_states_157")]; + tensor var_3241 = const()[name = string("op_3241"), val = tensor([1, 1, 8, 128])]; + tensor v_39 = reshape(shape = var_3241, x = v_37)[name = string("v_39")]; + fp32 var_3246_promoted = const()[name = string("op_3246_promoted"), val = fp32(0x1p+1)]; + tensor var_3252 = pow(x = hidden_states_151, y = var_3246_promoted)[name = string("op_3252")]; + tensor variance_51_axes_0 = const()[name = string("variance_51_axes_0"), val = tensor([-1])]; + bool variance_51_keep_dims_0 = const()[name = string("variance_51_keep_dims_0"), val = bool(true)]; + tensor variance_51 = reduce_mean(axes = variance_51_axes_0, keep_dims = variance_51_keep_dims_0, x = var_3252)[name = string("variance_51")]; + fp32 var_3255 = const()[name = string("op_3255"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3256 = add(x = variance_51, y = var_3255)[name = string("op_3256")]; + fp32 var_3257_epsilon_0 = const()[name = string("op_3257_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3257 = rsqrt(epsilon = var_3257_epsilon_0, x = var_3256)[name = string("op_3257")]; + tensor hidden_states_155 = mul(x = hidden_states_151, y = var_3257)[name = string("hidden_states_155")]; + tensor const_62 = const()[name = string("const_62"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774289344)))]; + tensor q_51 = mul(x = const_62, y = hidden_states_155)[name = string("q_51")]; + fp32 var_3264_promoted = const()[name = string("op_3264_promoted"), val = fp32(0x1p+1)]; + tensor var_3270 = pow(x = hidden_states_157, y = var_3264_promoted)[name = string("op_3270")]; + tensor variance_53_axes_0 = const()[name = string("variance_53_axes_0"), val = tensor([-1])]; + bool variance_53_keep_dims_0 = const()[name = string("variance_53_keep_dims_0"), val = bool(true)]; + tensor variance_53 = reduce_mean(axes = variance_53_axes_0, keep_dims = variance_53_keep_dims_0, x = var_3270)[name = string("variance_53")]; + fp32 var_3273 = const()[name = string("op_3273"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3274 = add(x = variance_53, y = var_3273)[name = string("op_3274")]; + fp32 var_3275_epsilon_0 = const()[name = string("op_3275_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3275 = rsqrt(epsilon = var_3275_epsilon_0, x = var_3274)[name = string("op_3275")]; + tensor hidden_states_161 = mul(x = hidden_states_157, y = var_3275)[name = string("hidden_states_161")]; + tensor const_63 = const()[name = string("const_63"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774289920)))]; + tensor k_51 = mul(x = const_63, y = hidden_states_161)[name = string("k_51")]; + tensor q_53_perm_0 = const()[name = string("q_53_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_53_perm_0 = const()[name = string("k_53_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_53 = transpose(perm = q_53_perm_0, x = q_51)[name = string("transpose_87")]; + tensor var_3292 = mul(x = q_53, y = cos_3)[name = string("op_3292")]; + tensor x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_25 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_53)[name = string("x1_25")]; + tensor x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_25 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_53)[name = string("x2_25")]; + fp32 const_66_promoted = const()[name = string("const_66_promoted"), val = fp32(-0x1p+0)]; + tensor var_3313 = mul(x = x2_25, y = const_66_promoted)[name = string("op_3313")]; + int32 var_3315 = const()[name = string("op_3315"), val = int32(-1)]; + bool var_3316_interleave_0 = const()[name = string("op_3316_interleave_0"), val = bool(false)]; + tensor var_3316 = concat(axis = var_3315, interleave = var_3316_interleave_0, values = (var_3313, x1_25))[name = string("op_3316")]; + tensor var_3317 = mul(x = var_3316, y = sin_3)[name = string("op_3317")]; + tensor q_55 = add(x = var_3292, y = var_3317)[name = string("q_55")]; + tensor k_53 = transpose(perm = k_53_perm_0, x = k_51)[name = string("transpose_86")]; + tensor var_3320 = mul(x = k_53, y = cos_3)[name = string("op_3320")]; + tensor x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_27 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_53)[name = string("x1_27")]; + tensor x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_27 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_53)[name = string("x2_27")]; + fp32 const_69_promoted = const()[name = string("const_69_promoted"), val = fp32(-0x1p+0)]; + tensor var_3341 = mul(x = x2_27, y = const_69_promoted)[name = string("op_3341")]; + int32 var_3343 = const()[name = string("op_3343"), val = int32(-1)]; + bool var_3344_interleave_0 = const()[name = string("op_3344_interleave_0"), val = bool(false)]; + tensor var_3344 = concat(axis = var_3343, interleave = var_3344_interleave_0, values = (var_3341, x1_27))[name = string("op_3344")]; + tensor var_3345 = mul(x = var_3344, y = sin_3)[name = string("op_3345")]; + tensor k_55 = add(x = var_3320, y = var_3345)[name = string("k_55")]; + tensor var_3352 = const()[name = string("op_3352"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_13 = reshape(shape = var_3352, x = k_55)[name = string("nk_flat_13")]; + tensor var_3358 = const()[name = string("op_3358"), val = tensor([1, 1024, 1, 1])]; + tensor v_41 = transpose(perm = v_41_perm_0, x = v_39)[name = string("transpose_85")]; + tensor nv_flat_13 = reshape(shape = var_3358, x = v_41)[name = string("nv_flat_13")]; + tensor var_3367 = mul(x = var_3161, y = var_1194)[name = string("op_3367")]; + tensor var_3368 = mul(x = nk_flat_13, y = update_mask_1)[name = string("op_3368")]; + tensor key_cache_29 = add(x = var_3367, y = var_3368)[name = string("key_cache_29")]; + tensor var_3374 = mul(x = var_3181, y = var_1194)[name = string("op_3374")]; + tensor var_3375 = mul(x = nv_flat_13, y = update_mask_1)[name = string("op_3375")]; + tensor value_cache_29 = add(x = var_3374, y = var_3375)[name = string("value_cache_29")]; + tensor kc_37_axes_0 = const()[name = string("kc_37_axes_0"), val = tensor([2])]; + tensor kc_37 = squeeze(axes = kc_37_axes_0, x = key_cache_29)[name = string("kc_37")]; + tensor var_3384 = const()[name = string("op_3384"), val = tensor([1, 8, 128, 256])]; + tensor kc_39 = reshape(shape = var_3384, x = kc_37)[name = string("kc_39")]; + tensor vc_37_axes_0 = const()[name = string("vc_37_axes_0"), val = tensor([2])]; + tensor vc_37 = squeeze(axes = vc_37_axes_0, x = value_cache_29)[name = string("vc_37")]; + tensor var_3392 = const()[name = string("op_3392"), val = tensor([1, 8, 128, 256])]; + tensor vc_39 = reshape(shape = var_3392, x = vc_37)[name = string("vc_39")]; + tensor var_3395_axes_0 = const()[name = string("op_3395_axes_0"), val = tensor([2])]; + tensor var_3395 = expand_dims(axes = var_3395_axes_0, x = kc_39)[name = string("op_3395")]; + tensor var_3403_reps_0 = const()[name = string("op_3403_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_3403 = tile(reps = var_3403_reps_0, x = var_3395)[name = string("op_3403")]; + tensor var_3408 = const()[name = string("op_3408"), val = tensor([1, 16, 128, 256])]; + tensor kc_41 = reshape(shape = var_3408, x = var_3403)[name = string("kc_41")]; + tensor var_3411_axes_0 = const()[name = string("op_3411_axes_0"), val = tensor([2])]; + tensor var_3411 = expand_dims(axes = var_3411_axes_0, x = vc_39)[name = string("op_3411")]; + tensor var_3419_reps_0 = const()[name = string("op_3419_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_3419 = tile(reps = var_3419_reps_0, x = var_3411)[name = string("op_3419")]; + tensor var_3424 = const()[name = string("op_3424"), val = tensor([1, 16, 128, 256])]; + tensor vc_41 = reshape(shape = var_3424, x = var_3419)[name = string("vc_41")]; + bool var_3426_transpose_x_0 = const()[name = string("op_3426_transpose_x_0"), val = bool(false)]; + bool var_3426_transpose_y_0 = const()[name = string("op_3426_transpose_y_0"), val = bool(false)]; + tensor var_3426 = matmul(transpose_x = var_3426_transpose_x_0, transpose_y = var_3426_transpose_y_0, x = q_55, y = kc_41)[name = string("op_3426")]; + fp32 _inversed_attn_weights_49_y_0 = const()[name = string("_inversed_attn_weights_49_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_49 = mul(x = var_3426, y = _inversed_attn_weights_49_y_0)[name = string("_inversed_attn_weights_49")]; + tensor attn_weights_51 = add(x = _inversed_attn_weights_49, y = mask_1)[name = string("attn_weights_51")]; + int32 var_3440 = const()[name = string("op_3440"), val = int32(-1)]; + tensor attn_weights_55 = softmax(axis = var_3440, x = attn_weights_51)[name = string("attn_weights_55")]; + bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(false)]; + bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(true)]; + tensor attn_output_25 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_55, y = vc_41)[name = string("attn_output_25")]; + tensor var_3449_perm_0 = const()[name = string("op_3449_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3453 = const()[name = string("op_3453"), val = tensor([1, 1, -1])]; + tensor var_3449 = transpose(perm = var_3449_perm_0, x = attn_output_25)[name = string("transpose_84")]; + tensor input_63 = reshape(shape = var_3453, x = var_3449)[name = string("input_63")]; + tensor attn_output_27 = linear(bias = linear_1_bias_0, weight = layers_6_self_attn_o_proj_weight, x = input_63)[name = string("linear_45")]; + tensor var_3459_axes_0 = const()[name = string("op_3459_axes_0"), val = tensor([0])]; + tensor var_3459 = squeeze(axes = var_3459_axes_0, x = attn_output_27)[name = string("op_3459")]; + tensor var_3461_axes_0 = const()[name = string("op_3461_axes_0"), val = tensor([0])]; + tensor var_3461 = squeeze(axes = var_3461_axes_0, x = var_3459)[name = string("op_3461")]; + tensor var_3463_axes_0 = const()[name = string("op_3463_axes_0"), val = tensor([-1])]; + tensor var_3463 = expand_dims(axes = var_3463_axes_0, x = var_3461)[name = string("op_3463")]; + tensor attn_4d_13_axes_0 = const()[name = string("attn_4d_13_axes_0"), val = tensor([-1])]; + tensor attn_4d_13 = expand_dims(axes = attn_4d_13_axes_0, x = var_3463)[name = string("attn_4d_13")]; + tensor hidden_25 = add(x = hidden_23, y = attn_4d_13)[name = string("hidden_25")]; + tensor var_3469_axes_0 = const()[name = string("op_3469_axes_0"), val = tensor([-1])]; + tensor var_3469 = squeeze(axes = var_3469_axes_0, x = hidden_25)[name = string("op_3469")]; + tensor var_3471_axes_0 = const()[name = string("op_3471_axes_0"), val = tensor([-1])]; + tensor var_3471 = squeeze(axes = var_3471_axes_0, x = var_3469)[name = string("op_3471")]; + tensor hidden_states_163_axes_0 = const()[name = string("hidden_states_163_axes_0"), val = tensor([0])]; + tensor hidden_states_163 = expand_dims(axes = hidden_states_163_axes_0, x = var_3471)[name = string("hidden_states_163")]; + fp32 var_3477_promoted = const()[name = string("op_3477_promoted"), val = fp32(0x1p+1)]; + tensor var_3483 = pow(x = hidden_states_163, y = var_3477_promoted)[name = string("op_3483")]; + tensor variance_55_axes_0 = const()[name = string("variance_55_axes_0"), val = tensor([-1])]; + bool variance_55_keep_dims_0 = const()[name = string("variance_55_keep_dims_0"), val = bool(true)]; + tensor variance_55 = reduce_mean(axes = variance_55_axes_0, keep_dims = variance_55_keep_dims_0, x = var_3483)[name = string("variance_55")]; + fp32 var_3486 = const()[name = string("op_3486"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3487 = add(x = variance_55, y = var_3486)[name = string("op_3487")]; + fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3488 = rsqrt(epsilon = var_3488_epsilon_0, x = var_3487)[name = string("op_3488")]; + tensor hidden_states_167 = mul(x = hidden_states_163, y = var_3488)[name = string("hidden_states_167")]; + tensor const_70 = const()[name = string("const_70"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774290496)))]; + tensor input_65 = mul(x = const_70, y = hidden_states_167)[name = string("input_65")]; + tensor input_67 = linear(bias = linear_4_bias_0, weight = layers_6_mlp_gate_proj_weight, x = input_65)[name = string("linear_46")]; + tensor var_3498 = silu(x = input_67)[name = string("op_3498")]; + tensor var_3500 = linear(bias = linear_4_bias_0, weight = layers_6_mlp_up_proj_weight, x = input_65)[name = string("linear_47")]; + tensor input_69 = mul(x = var_3498, y = var_3500)[name = string("input_69")]; + tensor mlp_out_13 = linear(bias = linear_1_bias_0, weight = layers_6_mlp_down_proj_weight, x = input_69)[name = string("linear_48")]; + tensor var_3505_axes_0 = const()[name = string("op_3505_axes_0"), val = tensor([0])]; + tensor var_3505 = squeeze(axes = var_3505_axes_0, x = mlp_out_13)[name = string("op_3505")]; + tensor var_3507_axes_0 = const()[name = string("op_3507_axes_0"), val = tensor([0])]; + tensor var_3507 = squeeze(axes = var_3507_axes_0, x = var_3505)[name = string("op_3507")]; + tensor var_3509_axes_0 = const()[name = string("op_3509_axes_0"), val = tensor([-1])]; + tensor var_3509 = expand_dims(axes = var_3509_axes_0, x = var_3507)[name = string("op_3509")]; + tensor mlp_4d_13_axes_0 = const()[name = string("mlp_4d_13_axes_0"), val = tensor([-1])]; + tensor mlp_4d_13 = expand_dims(axes = mlp_4d_13_axes_0, x = var_3509)[name = string("mlp_4d_13")]; + tensor hidden_27 = add(x = hidden_25, y = mlp_4d_13)[name = string("hidden_27")]; + tensor var_3523_begin_0 = const()[name = string("op_3523_begin_0"), val = tensor([0, 7168, 0, 0])]; + tensor var_3523_end_0 = const()[name = string("op_3523_end_0"), val = tensor([1, 8192, 1, 256])]; + tensor var_3523_end_mask_0 = const()[name = string("op_3523_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3523 = slice_by_index(begin = var_3523_begin_0, end = var_3523_end_0, end_mask = var_3523_end_mask_0, x = cast_3)[name = string("op_3523")]; + tensor var_3543_begin_0 = const()[name = string("op_3543_begin_0"), val = tensor([0, 7168, 0, 0])]; + tensor var_3543_end_0 = const()[name = string("op_3543_end_0"), val = tensor([1, 8192, 1, 256])]; + tensor var_3543_end_mask_0 = const()[name = string("op_3543_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3543 = slice_by_index(begin = var_3543_begin_0, end = var_3543_end_0, end_mask = var_3543_end_mask_0, x = cast_4)[name = string("op_3543")]; + tensor var_3555_axes_0 = const()[name = string("op_3555_axes_0"), val = tensor([-1])]; + tensor var_3555 = squeeze(axes = var_3555_axes_0, x = hidden_27)[name = string("op_3555")]; + tensor var_3557_axes_0 = const()[name = string("op_3557_axes_0"), val = tensor([-1])]; + tensor var_3557 = squeeze(axes = var_3557_axes_0, x = var_3555)[name = string("op_3557")]; + tensor hidden_states_169_axes_0 = const()[name = string("hidden_states_169_axes_0"), val = tensor([0])]; + tensor hidden_states_169 = expand_dims(axes = hidden_states_169_axes_0, x = var_3557)[name = string("hidden_states_169")]; + fp32 var_3563_promoted = const()[name = string("op_3563_promoted"), val = fp32(0x1p+1)]; + tensor var_3569 = pow(x = hidden_states_169, y = var_3563_promoted)[name = string("op_3569")]; + tensor variance_57_axes_0 = const()[name = string("variance_57_axes_0"), val = tensor([-1])]; + bool variance_57_keep_dims_0 = const()[name = string("variance_57_keep_dims_0"), val = bool(true)]; + tensor variance_57 = reduce_mean(axes = variance_57_axes_0, keep_dims = variance_57_keep_dims_0, x = var_3569)[name = string("variance_57")]; + fp32 var_3572 = const()[name = string("op_3572"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3573 = add(x = variance_57, y = var_3572)[name = string("op_3573")]; + fp32 var_3574_epsilon_0 = const()[name = string("op_3574_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3574 = rsqrt(epsilon = var_3574_epsilon_0, x = var_3573)[name = string("op_3574")]; + tensor hidden_states_173 = mul(x = hidden_states_169, y = var_3574)[name = string("hidden_states_173")]; + tensor const_71 = const()[name = string("const_71"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774294656)))]; + tensor input_71 = mul(x = const_71, y = hidden_states_173)[name = string("input_71")]; + tensor q_57 = linear(bias = linear_0_bias_0, weight = layers_7_self_attn_q_proj_weight, x = input_71)[name = string("linear_49")]; + tensor k_57 = linear(bias = linear_1_bias_0, weight = layers_7_self_attn_k_proj_weight, x = input_71)[name = string("linear_50")]; + tensor v_43 = linear(bias = linear_1_bias_0, weight = layers_7_self_attn_v_proj_weight, x = input_71)[name = string("linear_51")]; + tensor var_3591 = const()[name = string("op_3591"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_175 = reshape(shape = var_3591, x = q_57)[name = string("hidden_states_175")]; + tensor var_3597 = const()[name = string("op_3597"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_181 = reshape(shape = var_3597, x = k_57)[name = string("hidden_states_181")]; + tensor var_3603 = const()[name = string("op_3603"), val = tensor([1, 1, 8, 128])]; + tensor v_45 = reshape(shape = var_3603, x = v_43)[name = string("v_45")]; + fp32 var_3608_promoted = const()[name = string("op_3608_promoted"), val = fp32(0x1p+1)]; + tensor var_3614 = pow(x = hidden_states_175, y = var_3608_promoted)[name = string("op_3614")]; + tensor variance_59_axes_0 = const()[name = string("variance_59_axes_0"), val = tensor([-1])]; + bool variance_59_keep_dims_0 = const()[name = string("variance_59_keep_dims_0"), val = bool(true)]; + tensor variance_59 = reduce_mean(axes = variance_59_axes_0, keep_dims = variance_59_keep_dims_0, x = var_3614)[name = string("variance_59")]; + fp32 var_3617 = const()[name = string("op_3617"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3618 = add(x = variance_59, y = var_3617)[name = string("op_3618")]; + fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3619 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618)[name = string("op_3619")]; + tensor hidden_states_179 = mul(x = hidden_states_175, y = var_3619)[name = string("hidden_states_179")]; + tensor const_72 = const()[name = string("const_72"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774298816)))]; + tensor q_59 = mul(x = const_72, y = hidden_states_179)[name = string("q_59")]; + fp32 var_3626_promoted = const()[name = string("op_3626_promoted"), val = fp32(0x1p+1)]; + tensor var_3632 = pow(x = hidden_states_181, y = var_3626_promoted)[name = string("op_3632")]; + tensor variance_61_axes_0 = const()[name = string("variance_61_axes_0"), val = tensor([-1])]; + bool variance_61_keep_dims_0 = const()[name = string("variance_61_keep_dims_0"), val = bool(true)]; + tensor variance_61 = reduce_mean(axes = variance_61_axes_0, keep_dims = variance_61_keep_dims_0, x = var_3632)[name = string("variance_61")]; + fp32 var_3635 = const()[name = string("op_3635"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3636 = add(x = variance_61, y = var_3635)[name = string("op_3636")]; + fp32 var_3637_epsilon_0 = const()[name = string("op_3637_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3637 = rsqrt(epsilon = var_3637_epsilon_0, x = var_3636)[name = string("op_3637")]; + tensor hidden_states_185 = mul(x = hidden_states_181, y = var_3637)[name = string("hidden_states_185")]; + tensor const_73 = const()[name = string("const_73"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774299392)))]; + tensor k_59 = mul(x = const_73, y = hidden_states_185)[name = string("k_59")]; + tensor q_61_perm_0 = const()[name = string("q_61_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_61_perm_0 = const()[name = string("k_61_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_61 = transpose(perm = q_61_perm_0, x = q_59)[name = string("transpose_83")]; + tensor var_3654 = mul(x = q_61, y = cos_3)[name = string("op_3654")]; + tensor x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_29 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_61)[name = string("x1_29")]; + tensor x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_29 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_61)[name = string("x2_29")]; + fp32 const_76_promoted = const()[name = string("const_76_promoted"), val = fp32(-0x1p+0)]; + tensor var_3675 = mul(x = x2_29, y = const_76_promoted)[name = string("op_3675")]; + int32 var_3677 = const()[name = string("op_3677"), val = int32(-1)]; + bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)]; + tensor var_3678 = concat(axis = var_3677, interleave = var_3678_interleave_0, values = (var_3675, x1_29))[name = string("op_3678")]; + tensor var_3679 = mul(x = var_3678, y = sin_3)[name = string("op_3679")]; + tensor q_63 = add(x = var_3654, y = var_3679)[name = string("q_63")]; + tensor k_61 = transpose(perm = k_61_perm_0, x = k_59)[name = string("transpose_82")]; + tensor var_3682 = mul(x = k_61, y = cos_3)[name = string("op_3682")]; + tensor x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_31 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_61)[name = string("x1_31")]; + tensor x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_31 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_61)[name = string("x2_31")]; + fp32 const_79_promoted = const()[name = string("const_79_promoted"), val = fp32(-0x1p+0)]; + tensor var_3703 = mul(x = x2_31, y = const_79_promoted)[name = string("op_3703")]; + int32 var_3705 = const()[name = string("op_3705"), val = int32(-1)]; + bool var_3706_interleave_0 = const()[name = string("op_3706_interleave_0"), val = bool(false)]; + tensor var_3706 = concat(axis = var_3705, interleave = var_3706_interleave_0, values = (var_3703, x1_31))[name = string("op_3706")]; + tensor var_3707 = mul(x = var_3706, y = sin_3)[name = string("op_3707")]; + tensor k_63 = add(x = var_3682, y = var_3707)[name = string("k_63")]; + tensor var_3714 = const()[name = string("op_3714"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_15 = reshape(shape = var_3714, x = k_63)[name = string("nk_flat_15")]; + tensor var_3720 = const()[name = string("op_3720"), val = tensor([1, 1024, 1, 1])]; + tensor v_47 = transpose(perm = v_47_perm_0, x = v_45)[name = string("transpose_81")]; + tensor nv_flat_15 = reshape(shape = var_3720, x = v_47)[name = string("nv_flat_15")]; + tensor var_3729 = mul(x = var_3523, y = var_1194)[name = string("op_3729")]; + tensor var_3730 = mul(x = nk_flat_15, y = update_mask_1)[name = string("op_3730")]; + tensor key_cache_33 = add(x = var_3729, y = var_3730)[name = string("key_cache_33")]; + tensor var_3736 = mul(x = var_3543, y = var_1194)[name = string("op_3736")]; + tensor var_3737 = mul(x = nv_flat_15, y = update_mask_1)[name = string("op_3737")]; + tensor value_cache_33 = add(x = var_3736, y = var_3737)[name = string("value_cache_33")]; + tensor kc_43_axes_0 = const()[name = string("kc_43_axes_0"), val = tensor([2])]; + tensor kc_43 = squeeze(axes = kc_43_axes_0, x = key_cache_33)[name = string("kc_43")]; + tensor var_3746 = const()[name = string("op_3746"), val = tensor([1, 8, 128, 256])]; + tensor kc_45 = reshape(shape = var_3746, x = kc_43)[name = string("kc_45")]; + tensor vc_43_axes_0 = const()[name = string("vc_43_axes_0"), val = tensor([2])]; + tensor vc_43 = squeeze(axes = vc_43_axes_0, x = value_cache_33)[name = string("vc_43")]; + tensor var_3754 = const()[name = string("op_3754"), val = tensor([1, 8, 128, 256])]; + tensor vc_45 = reshape(shape = var_3754, x = vc_43)[name = string("vc_45")]; + tensor var_3757_axes_0 = const()[name = string("op_3757_axes_0"), val = tensor([2])]; + tensor var_3757 = expand_dims(axes = var_3757_axes_0, x = kc_45)[name = string("op_3757")]; + tensor var_3765_reps_0 = const()[name = string("op_3765_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_3765 = tile(reps = var_3765_reps_0, x = var_3757)[name = string("op_3765")]; + tensor var_3770 = const()[name = string("op_3770"), val = tensor([1, 16, 128, 256])]; + tensor kc_47 = reshape(shape = var_3770, x = var_3765)[name = string("kc_47")]; + tensor var_3773_axes_0 = const()[name = string("op_3773_axes_0"), val = tensor([2])]; + tensor var_3773 = expand_dims(axes = var_3773_axes_0, x = vc_45)[name = string("op_3773")]; + tensor var_3781_reps_0 = const()[name = string("op_3781_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_3781 = tile(reps = var_3781_reps_0, x = var_3773)[name = string("op_3781")]; + tensor var_3786 = const()[name = string("op_3786"), val = tensor([1, 16, 128, 256])]; + tensor vc_47 = reshape(shape = var_3786, x = var_3781)[name = string("vc_47")]; + bool var_3788_transpose_x_0 = const()[name = string("op_3788_transpose_x_0"), val = bool(false)]; + bool var_3788_transpose_y_0 = const()[name = string("op_3788_transpose_y_0"), val = bool(false)]; + tensor var_3788 = matmul(transpose_x = var_3788_transpose_x_0, transpose_y = var_3788_transpose_y_0, x = q_63, y = kc_47)[name = string("op_3788")]; + fp32 _inversed_attn_weights_57_y_0 = const()[name = string("_inversed_attn_weights_57_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_57 = mul(x = var_3788, y = _inversed_attn_weights_57_y_0)[name = string("_inversed_attn_weights_57")]; + tensor attn_weights_59 = add(x = _inversed_attn_weights_57, y = mask_1)[name = string("attn_weights_59")]; + int32 var_3802 = const()[name = string("op_3802"), val = int32(-1)]; + tensor attn_weights_63 = softmax(axis = var_3802, x = attn_weights_59)[name = string("attn_weights_63")]; + bool attn_output_29_transpose_x_1 = const()[name = string("attn_output_29_transpose_x_1"), val = bool(false)]; + bool attn_output_29_transpose_y_1 = const()[name = string("attn_output_29_transpose_y_1"), val = bool(true)]; + tensor attn_output_29 = matmul(transpose_x = attn_output_29_transpose_x_1, transpose_y = attn_output_29_transpose_y_1, x = attn_weights_63, y = vc_47)[name = string("attn_output_29")]; + tensor var_3811_perm_0 = const()[name = string("op_3811_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3815 = const()[name = string("op_3815"), val = tensor([1, 1, -1])]; + tensor var_3811 = transpose(perm = var_3811_perm_0, x = attn_output_29)[name = string("transpose_80")]; + tensor input_73 = reshape(shape = var_3815, x = var_3811)[name = string("input_73")]; + tensor attn_output_31 = linear(bias = linear_1_bias_0, weight = layers_7_self_attn_o_proj_weight, x = input_73)[name = string("linear_52")]; + tensor var_3821_axes_0 = const()[name = string("op_3821_axes_0"), val = tensor([0])]; + tensor var_3821 = squeeze(axes = var_3821_axes_0, x = attn_output_31)[name = string("op_3821")]; + tensor var_3823_axes_0 = const()[name = string("op_3823_axes_0"), val = tensor([0])]; + tensor var_3823 = squeeze(axes = var_3823_axes_0, x = var_3821)[name = string("op_3823")]; + tensor var_3825_axes_0 = const()[name = string("op_3825_axes_0"), val = tensor([-1])]; + tensor var_3825 = expand_dims(axes = var_3825_axes_0, x = var_3823)[name = string("op_3825")]; + tensor attn_4d_15_axes_0 = const()[name = string("attn_4d_15_axes_0"), val = tensor([-1])]; + tensor attn_4d_15 = expand_dims(axes = attn_4d_15_axes_0, x = var_3825)[name = string("attn_4d_15")]; + tensor hidden_29 = add(x = hidden_27, y = attn_4d_15)[name = string("hidden_29")]; + tensor var_3831_axes_0 = const()[name = string("op_3831_axes_0"), val = tensor([-1])]; + tensor var_3831 = squeeze(axes = var_3831_axes_0, x = hidden_29)[name = string("op_3831")]; + tensor var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor([-1])]; + tensor var_3833 = squeeze(axes = var_3833_axes_0, x = var_3831)[name = string("op_3833")]; + tensor hidden_states_187_axes_0 = const()[name = string("hidden_states_187_axes_0"), val = tensor([0])]; + tensor hidden_states_187 = expand_dims(axes = hidden_states_187_axes_0, x = var_3833)[name = string("hidden_states_187")]; + fp32 var_3839_promoted = const()[name = string("op_3839_promoted"), val = fp32(0x1p+1)]; + tensor var_3845 = pow(x = hidden_states_187, y = var_3839_promoted)[name = string("op_3845")]; + tensor variance_63_axes_0 = const()[name = string("variance_63_axes_0"), val = tensor([-1])]; + bool variance_63_keep_dims_0 = const()[name = string("variance_63_keep_dims_0"), val = bool(true)]; + tensor variance_63 = reduce_mean(axes = variance_63_axes_0, keep_dims = variance_63_keep_dims_0, x = var_3845)[name = string("variance_63")]; + fp32 var_3848 = const()[name = string("op_3848"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3849 = add(x = variance_63, y = var_3848)[name = string("op_3849")]; + fp32 var_3850_epsilon_0 = const()[name = string("op_3850_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3850 = rsqrt(epsilon = var_3850_epsilon_0, x = var_3849)[name = string("op_3850")]; + tensor hidden_states_191 = mul(x = hidden_states_187, y = var_3850)[name = string("hidden_states_191")]; + tensor const_80 = const()[name = string("const_80"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774299968)))]; + tensor input_75 = mul(x = const_80, y = hidden_states_191)[name = string("input_75")]; + tensor input_77 = linear(bias = linear_4_bias_0, weight = layers_7_mlp_gate_proj_weight, x = input_75)[name = string("linear_53")]; + tensor var_3860 = silu(x = input_77)[name = string("op_3860")]; + tensor var_3862 = linear(bias = linear_4_bias_0, weight = layers_7_mlp_up_proj_weight, x = input_75)[name = string("linear_54")]; + tensor input_79 = mul(x = var_3860, y = var_3862)[name = string("input_79")]; + tensor mlp_out_15 = linear(bias = linear_1_bias_0, weight = layers_7_mlp_down_proj_weight, x = input_79)[name = string("linear_55")]; + tensor var_3867_axes_0 = const()[name = string("op_3867_axes_0"), val = tensor([0])]; + tensor var_3867 = squeeze(axes = var_3867_axes_0, x = mlp_out_15)[name = string("op_3867")]; + tensor var_3869_axes_0 = const()[name = string("op_3869_axes_0"), val = tensor([0])]; + tensor var_3869 = squeeze(axes = var_3869_axes_0, x = var_3867)[name = string("op_3869")]; + tensor var_3871_axes_0 = const()[name = string("op_3871_axes_0"), val = tensor([-1])]; + tensor var_3871 = expand_dims(axes = var_3871_axes_0, x = var_3869)[name = string("op_3871")]; + tensor mlp_4d_15_axes_0 = const()[name = string("mlp_4d_15_axes_0"), val = tensor([-1])]; + tensor mlp_4d_15 = expand_dims(axes = mlp_4d_15_axes_0, x = var_3871)[name = string("mlp_4d_15")]; + tensor hidden_31 = add(x = hidden_29, y = mlp_4d_15)[name = string("hidden_31")]; + tensor var_3885_begin_0 = const()[name = string("op_3885_begin_0"), val = tensor([0, 8192, 0, 0])]; + tensor var_3885_end_0 = const()[name = string("op_3885_end_0"), val = tensor([1, 9216, 1, 256])]; + tensor var_3885_end_mask_0 = const()[name = string("op_3885_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3885 = slice_by_index(begin = var_3885_begin_0, end = var_3885_end_0, end_mask = var_3885_end_mask_0, x = cast_3)[name = string("op_3885")]; + tensor var_3905_begin_0 = const()[name = string("op_3905_begin_0"), val = tensor([0, 8192, 0, 0])]; + tensor var_3905_end_0 = const()[name = string("op_3905_end_0"), val = tensor([1, 9216, 1, 256])]; + tensor var_3905_end_mask_0 = const()[name = string("op_3905_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_3905 = slice_by_index(begin = var_3905_begin_0, end = var_3905_end_0, end_mask = var_3905_end_mask_0, x = cast_4)[name = string("op_3905")]; + tensor var_3917_axes_0 = const()[name = string("op_3917_axes_0"), val = tensor([-1])]; + tensor var_3917 = squeeze(axes = var_3917_axes_0, x = hidden_31)[name = string("op_3917")]; + tensor var_3919_axes_0 = const()[name = string("op_3919_axes_0"), val = tensor([-1])]; + tensor var_3919 = squeeze(axes = var_3919_axes_0, x = var_3917)[name = string("op_3919")]; + tensor hidden_states_193_axes_0 = const()[name = string("hidden_states_193_axes_0"), val = tensor([0])]; + tensor hidden_states_193 = expand_dims(axes = hidden_states_193_axes_0, x = var_3919)[name = string("hidden_states_193")]; + fp32 var_3925_promoted = const()[name = string("op_3925_promoted"), val = fp32(0x1p+1)]; + tensor var_3931 = pow(x = hidden_states_193, y = var_3925_promoted)[name = string("op_3931")]; + tensor variance_65_axes_0 = const()[name = string("variance_65_axes_0"), val = tensor([-1])]; + bool variance_65_keep_dims_0 = const()[name = string("variance_65_keep_dims_0"), val = bool(true)]; + tensor variance_65 = reduce_mean(axes = variance_65_axes_0, keep_dims = variance_65_keep_dims_0, x = var_3931)[name = string("variance_65")]; + fp32 var_3934 = const()[name = string("op_3934"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3935 = add(x = variance_65, y = var_3934)[name = string("op_3935")]; + fp32 var_3936_epsilon_0 = const()[name = string("op_3936_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3936 = rsqrt(epsilon = var_3936_epsilon_0, x = var_3935)[name = string("op_3936")]; + tensor hidden_states_197 = mul(x = hidden_states_193, y = var_3936)[name = string("hidden_states_197")]; + tensor const_81 = const()[name = string("const_81"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774304128)))]; + tensor input_81 = mul(x = const_81, y = hidden_states_197)[name = string("input_81")]; + tensor q_65 = linear(bias = linear_0_bias_0, weight = layers_8_self_attn_q_proj_weight, x = input_81)[name = string("linear_56")]; + tensor k_65 = linear(bias = linear_1_bias_0, weight = layers_8_self_attn_k_proj_weight, x = input_81)[name = string("linear_57")]; + tensor v_49 = linear(bias = linear_1_bias_0, weight = layers_8_self_attn_v_proj_weight, x = input_81)[name = string("linear_58")]; + tensor var_3953 = const()[name = string("op_3953"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_199 = reshape(shape = var_3953, x = q_65)[name = string("hidden_states_199")]; + tensor var_3959 = const()[name = string("op_3959"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_205 = reshape(shape = var_3959, x = k_65)[name = string("hidden_states_205")]; + tensor var_3965 = const()[name = string("op_3965"), val = tensor([1, 1, 8, 128])]; + tensor v_51 = reshape(shape = var_3965, x = v_49)[name = string("v_51")]; + fp32 var_3970_promoted = const()[name = string("op_3970_promoted"), val = fp32(0x1p+1)]; + tensor var_3976 = pow(x = hidden_states_199, y = var_3970_promoted)[name = string("op_3976")]; + tensor variance_67_axes_0 = const()[name = string("variance_67_axes_0"), val = tensor([-1])]; + bool variance_67_keep_dims_0 = const()[name = string("variance_67_keep_dims_0"), val = bool(true)]; + tensor variance_67 = reduce_mean(axes = variance_67_axes_0, keep_dims = variance_67_keep_dims_0, x = var_3976)[name = string("variance_67")]; + fp32 var_3979 = const()[name = string("op_3979"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3980 = add(x = variance_67, y = var_3979)[name = string("op_3980")]; + fp32 var_3981_epsilon_0 = const()[name = string("op_3981_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3981 = rsqrt(epsilon = var_3981_epsilon_0, x = var_3980)[name = string("op_3981")]; + tensor hidden_states_203 = mul(x = hidden_states_199, y = var_3981)[name = string("hidden_states_203")]; + tensor const_82 = const()[name = string("const_82"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774308288)))]; + tensor q_67 = mul(x = const_82, y = hidden_states_203)[name = string("q_67")]; + fp32 var_3988_promoted = const()[name = string("op_3988_promoted"), val = fp32(0x1p+1)]; + tensor var_3994 = pow(x = hidden_states_205, y = var_3988_promoted)[name = string("op_3994")]; + tensor variance_69_axes_0 = const()[name = string("variance_69_axes_0"), val = tensor([-1])]; + bool variance_69_keep_dims_0 = const()[name = string("variance_69_keep_dims_0"), val = bool(true)]; + tensor variance_69 = reduce_mean(axes = variance_69_axes_0, keep_dims = variance_69_keep_dims_0, x = var_3994)[name = string("variance_69")]; + fp32 var_3997 = const()[name = string("op_3997"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_3998 = add(x = variance_69, y = var_3997)[name = string("op_3998")]; + fp32 var_3999_epsilon_0 = const()[name = string("op_3999_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_3999 = rsqrt(epsilon = var_3999_epsilon_0, x = var_3998)[name = string("op_3999")]; + tensor hidden_states_209 = mul(x = hidden_states_205, y = var_3999)[name = string("hidden_states_209")]; + tensor const_83 = const()[name = string("const_83"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774308864)))]; + tensor k_67 = mul(x = const_83, y = hidden_states_209)[name = string("k_67")]; + tensor q_69_perm_0 = const()[name = string("q_69_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_69_perm_0 = const()[name = string("k_69_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_53_perm_0 = const()[name = string("v_53_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_69 = transpose(perm = q_69_perm_0, x = q_67)[name = string("transpose_79")]; + tensor var_4016 = mul(x = q_69, y = cos_3)[name = string("op_4016")]; + tensor x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_33 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_69)[name = string("x1_33")]; + tensor x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_33 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_69)[name = string("x2_33")]; + fp32 const_86_promoted = const()[name = string("const_86_promoted"), val = fp32(-0x1p+0)]; + tensor var_4037 = mul(x = x2_33, y = const_86_promoted)[name = string("op_4037")]; + int32 var_4039 = const()[name = string("op_4039"), val = int32(-1)]; + bool var_4040_interleave_0 = const()[name = string("op_4040_interleave_0"), val = bool(false)]; + tensor var_4040 = concat(axis = var_4039, interleave = var_4040_interleave_0, values = (var_4037, x1_33))[name = string("op_4040")]; + tensor var_4041 = mul(x = var_4040, y = sin_3)[name = string("op_4041")]; + tensor q_71 = add(x = var_4016, y = var_4041)[name = string("q_71")]; + tensor k_69 = transpose(perm = k_69_perm_0, x = k_67)[name = string("transpose_78")]; + tensor var_4044 = mul(x = k_69, y = cos_3)[name = string("op_4044")]; + tensor x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_35 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_69)[name = string("x1_35")]; + tensor x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_35 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_69)[name = string("x2_35")]; + fp32 const_89_promoted = const()[name = string("const_89_promoted"), val = fp32(-0x1p+0)]; + tensor var_4065 = mul(x = x2_35, y = const_89_promoted)[name = string("op_4065")]; + int32 var_4067 = const()[name = string("op_4067"), val = int32(-1)]; + bool var_4068_interleave_0 = const()[name = string("op_4068_interleave_0"), val = bool(false)]; + tensor var_4068 = concat(axis = var_4067, interleave = var_4068_interleave_0, values = (var_4065, x1_35))[name = string("op_4068")]; + tensor var_4069 = mul(x = var_4068, y = sin_3)[name = string("op_4069")]; + tensor k_71 = add(x = var_4044, y = var_4069)[name = string("k_71")]; + tensor var_4076 = const()[name = string("op_4076"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_17 = reshape(shape = var_4076, x = k_71)[name = string("nk_flat_17")]; + tensor var_4082 = const()[name = string("op_4082"), val = tensor([1, 1024, 1, 1])]; + tensor v_53 = transpose(perm = v_53_perm_0, x = v_51)[name = string("transpose_77")]; + tensor nv_flat_17 = reshape(shape = var_4082, x = v_53)[name = string("nv_flat_17")]; + tensor var_4091 = mul(x = var_3885, y = var_1194)[name = string("op_4091")]; + tensor var_4092 = mul(x = nk_flat_17, y = update_mask_1)[name = string("op_4092")]; + tensor key_cache_37 = add(x = var_4091, y = var_4092)[name = string("key_cache_37")]; + tensor var_4098 = mul(x = var_3905, y = var_1194)[name = string("op_4098")]; + tensor var_4099 = mul(x = nv_flat_17, y = update_mask_1)[name = string("op_4099")]; + tensor value_cache_37 = add(x = var_4098, y = var_4099)[name = string("value_cache_37")]; + tensor kc_49_axes_0 = const()[name = string("kc_49_axes_0"), val = tensor([2])]; + tensor kc_49 = squeeze(axes = kc_49_axes_0, x = key_cache_37)[name = string("kc_49")]; + tensor var_4108 = const()[name = string("op_4108"), val = tensor([1, 8, 128, 256])]; + tensor kc_51 = reshape(shape = var_4108, x = kc_49)[name = string("kc_51")]; + tensor vc_49_axes_0 = const()[name = string("vc_49_axes_0"), val = tensor([2])]; + tensor vc_49 = squeeze(axes = vc_49_axes_0, x = value_cache_37)[name = string("vc_49")]; + tensor var_4116 = const()[name = string("op_4116"), val = tensor([1, 8, 128, 256])]; + tensor vc_51 = reshape(shape = var_4116, x = vc_49)[name = string("vc_51")]; + tensor var_4119_axes_0 = const()[name = string("op_4119_axes_0"), val = tensor([2])]; + tensor var_4119 = expand_dims(axes = var_4119_axes_0, x = kc_51)[name = string("op_4119")]; + tensor var_4127_reps_0 = const()[name = string("op_4127_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_4127 = tile(reps = var_4127_reps_0, x = var_4119)[name = string("op_4127")]; + tensor var_4132 = const()[name = string("op_4132"), val = tensor([1, 16, 128, 256])]; + tensor kc_53 = reshape(shape = var_4132, x = var_4127)[name = string("kc_53")]; + tensor var_4135_axes_0 = const()[name = string("op_4135_axes_0"), val = tensor([2])]; + tensor var_4135 = expand_dims(axes = var_4135_axes_0, x = vc_51)[name = string("op_4135")]; + tensor var_4143_reps_0 = const()[name = string("op_4143_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_4143 = tile(reps = var_4143_reps_0, x = var_4135)[name = string("op_4143")]; + tensor var_4148 = const()[name = string("op_4148"), val = tensor([1, 16, 128, 256])]; + tensor vc_53 = reshape(shape = var_4148, x = var_4143)[name = string("vc_53")]; + bool var_4150_transpose_x_0 = const()[name = string("op_4150_transpose_x_0"), val = bool(false)]; + bool var_4150_transpose_y_0 = const()[name = string("op_4150_transpose_y_0"), val = bool(false)]; + tensor var_4150 = matmul(transpose_x = var_4150_transpose_x_0, transpose_y = var_4150_transpose_y_0, x = q_71, y = kc_53)[name = string("op_4150")]; + fp32 _inversed_attn_weights_65_y_0 = const()[name = string("_inversed_attn_weights_65_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_65 = mul(x = var_4150, y = _inversed_attn_weights_65_y_0)[name = string("_inversed_attn_weights_65")]; + tensor attn_weights_67 = add(x = _inversed_attn_weights_65, y = mask_1)[name = string("attn_weights_67")]; + int32 var_4164 = const()[name = string("op_4164"), val = int32(-1)]; + tensor attn_weights_71 = softmax(axis = var_4164, x = attn_weights_67)[name = string("attn_weights_71")]; + bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(false)]; + bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(true)]; + tensor attn_output_33 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_71, y = vc_53)[name = string("attn_output_33")]; + tensor var_4173_perm_0 = const()[name = string("op_4173_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4177 = const()[name = string("op_4177"), val = tensor([1, 1, -1])]; + tensor var_4173 = transpose(perm = var_4173_perm_0, x = attn_output_33)[name = string("transpose_76")]; + tensor input_83 = reshape(shape = var_4177, x = var_4173)[name = string("input_83")]; + tensor attn_output_35 = linear(bias = linear_1_bias_0, weight = layers_8_self_attn_o_proj_weight, x = input_83)[name = string("linear_59")]; + tensor var_4183_axes_0 = const()[name = string("op_4183_axes_0"), val = tensor([0])]; + tensor var_4183 = squeeze(axes = var_4183_axes_0, x = attn_output_35)[name = string("op_4183")]; + tensor var_4185_axes_0 = const()[name = string("op_4185_axes_0"), val = tensor([0])]; + tensor var_4185 = squeeze(axes = var_4185_axes_0, x = var_4183)[name = string("op_4185")]; + tensor var_4187_axes_0 = const()[name = string("op_4187_axes_0"), val = tensor([-1])]; + tensor var_4187 = expand_dims(axes = var_4187_axes_0, x = var_4185)[name = string("op_4187")]; + tensor attn_4d_17_axes_0 = const()[name = string("attn_4d_17_axes_0"), val = tensor([-1])]; + tensor attn_4d_17 = expand_dims(axes = attn_4d_17_axes_0, x = var_4187)[name = string("attn_4d_17")]; + tensor hidden_33 = add(x = hidden_31, y = attn_4d_17)[name = string("hidden_33")]; + tensor var_4193_axes_0 = const()[name = string("op_4193_axes_0"), val = tensor([-1])]; + tensor var_4193 = squeeze(axes = var_4193_axes_0, x = hidden_33)[name = string("op_4193")]; + tensor var_4195_axes_0 = const()[name = string("op_4195_axes_0"), val = tensor([-1])]; + tensor var_4195 = squeeze(axes = var_4195_axes_0, x = var_4193)[name = string("op_4195")]; + tensor hidden_states_211_axes_0 = const()[name = string("hidden_states_211_axes_0"), val = tensor([0])]; + tensor hidden_states_211 = expand_dims(axes = hidden_states_211_axes_0, x = var_4195)[name = string("hidden_states_211")]; + fp32 var_4201_promoted = const()[name = string("op_4201_promoted"), val = fp32(0x1p+1)]; + tensor var_4207 = pow(x = hidden_states_211, y = var_4201_promoted)[name = string("op_4207")]; + tensor variance_71_axes_0 = const()[name = string("variance_71_axes_0"), val = tensor([-1])]; + bool variance_71_keep_dims_0 = const()[name = string("variance_71_keep_dims_0"), val = bool(true)]; + tensor variance_71 = reduce_mean(axes = variance_71_axes_0, keep_dims = variance_71_keep_dims_0, x = var_4207)[name = string("variance_71")]; + fp32 var_4210 = const()[name = string("op_4210"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4211 = add(x = variance_71, y = var_4210)[name = string("op_4211")]; + fp32 var_4212_epsilon_0 = const()[name = string("op_4212_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4212 = rsqrt(epsilon = var_4212_epsilon_0, x = var_4211)[name = string("op_4212")]; + tensor hidden_states_215 = mul(x = hidden_states_211, y = var_4212)[name = string("hidden_states_215")]; + tensor const_90 = const()[name = string("const_90"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774309440)))]; + tensor input_85 = mul(x = const_90, y = hidden_states_215)[name = string("input_85")]; + tensor input_87 = linear(bias = linear_4_bias_0, weight = layers_8_mlp_gate_proj_weight, x = input_85)[name = string("linear_60")]; + tensor var_4222 = silu(x = input_87)[name = string("op_4222")]; + tensor var_4224 = linear(bias = linear_4_bias_0, weight = layers_8_mlp_up_proj_weight, x = input_85)[name = string("linear_61")]; + tensor input_89 = mul(x = var_4222, y = var_4224)[name = string("input_89")]; + tensor mlp_out_17 = linear(bias = linear_1_bias_0, weight = layers_8_mlp_down_proj_weight, x = input_89)[name = string("linear_62")]; + tensor var_4229_axes_0 = const()[name = string("op_4229_axes_0"), val = tensor([0])]; + tensor var_4229 = squeeze(axes = var_4229_axes_0, x = mlp_out_17)[name = string("op_4229")]; + tensor var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor([0])]; + tensor var_4231 = squeeze(axes = var_4231_axes_0, x = var_4229)[name = string("op_4231")]; + tensor var_4233_axes_0 = const()[name = string("op_4233_axes_0"), val = tensor([-1])]; + tensor var_4233 = expand_dims(axes = var_4233_axes_0, x = var_4231)[name = string("op_4233")]; + tensor mlp_4d_17_axes_0 = const()[name = string("mlp_4d_17_axes_0"), val = tensor([-1])]; + tensor mlp_4d_17 = expand_dims(axes = mlp_4d_17_axes_0, x = var_4233)[name = string("mlp_4d_17")]; + tensor hidden_35 = add(x = hidden_33, y = mlp_4d_17)[name = string("hidden_35")]; + tensor var_4247_begin_0 = const()[name = string("op_4247_begin_0"), val = tensor([0, 9216, 0, 0])]; + tensor var_4247_end_0 = const()[name = string("op_4247_end_0"), val = tensor([1, 10240, 1, 256])]; + tensor var_4247_end_mask_0 = const()[name = string("op_4247_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4247 = slice_by_index(begin = var_4247_begin_0, end = var_4247_end_0, end_mask = var_4247_end_mask_0, x = cast_3)[name = string("op_4247")]; + tensor var_4267_begin_0 = const()[name = string("op_4267_begin_0"), val = tensor([0, 9216, 0, 0])]; + tensor var_4267_end_0 = const()[name = string("op_4267_end_0"), val = tensor([1, 10240, 1, 256])]; + tensor var_4267_end_mask_0 = const()[name = string("op_4267_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4267 = slice_by_index(begin = var_4267_begin_0, end = var_4267_end_0, end_mask = var_4267_end_mask_0, x = cast_4)[name = string("op_4267")]; + tensor var_4279_axes_0 = const()[name = string("op_4279_axes_0"), val = tensor([-1])]; + tensor var_4279 = squeeze(axes = var_4279_axes_0, x = hidden_35)[name = string("op_4279")]; + tensor var_4281_axes_0 = const()[name = string("op_4281_axes_0"), val = tensor([-1])]; + tensor var_4281 = squeeze(axes = var_4281_axes_0, x = var_4279)[name = string("op_4281")]; + tensor hidden_states_217_axes_0 = const()[name = string("hidden_states_217_axes_0"), val = tensor([0])]; + tensor hidden_states_217 = expand_dims(axes = hidden_states_217_axes_0, x = var_4281)[name = string("hidden_states_217")]; + fp32 var_4287_promoted = const()[name = string("op_4287_promoted"), val = fp32(0x1p+1)]; + tensor var_4293 = pow(x = hidden_states_217, y = var_4287_promoted)[name = string("op_4293")]; + tensor variance_73_axes_0 = const()[name = string("variance_73_axes_0"), val = tensor([-1])]; + bool variance_73_keep_dims_0 = const()[name = string("variance_73_keep_dims_0"), val = bool(true)]; + tensor variance_73 = reduce_mean(axes = variance_73_axes_0, keep_dims = variance_73_keep_dims_0, x = var_4293)[name = string("variance_73")]; + fp32 var_4296 = const()[name = string("op_4296"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4297 = add(x = variance_73, y = var_4296)[name = string("op_4297")]; + fp32 var_4298_epsilon_0 = const()[name = string("op_4298_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4298 = rsqrt(epsilon = var_4298_epsilon_0, x = var_4297)[name = string("op_4298")]; + tensor hidden_states_221 = mul(x = hidden_states_217, y = var_4298)[name = string("hidden_states_221")]; + tensor const_91 = const()[name = string("const_91"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774313600)))]; + tensor input_91 = mul(x = const_91, y = hidden_states_221)[name = string("input_91")]; + tensor q_73 = linear(bias = linear_0_bias_0, weight = layers_9_self_attn_q_proj_weight, x = input_91)[name = string("linear_63")]; + tensor k_73 = linear(bias = linear_1_bias_0, weight = layers_9_self_attn_k_proj_weight, x = input_91)[name = string("linear_64")]; + tensor v_55 = linear(bias = linear_1_bias_0, weight = layers_9_self_attn_v_proj_weight, x = input_91)[name = string("linear_65")]; + tensor var_4315 = const()[name = string("op_4315"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_223 = reshape(shape = var_4315, x = q_73)[name = string("hidden_states_223")]; + tensor var_4321 = const()[name = string("op_4321"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_229 = reshape(shape = var_4321, x = k_73)[name = string("hidden_states_229")]; + tensor var_4327 = const()[name = string("op_4327"), val = tensor([1, 1, 8, 128])]; + tensor v_57 = reshape(shape = var_4327, x = v_55)[name = string("v_57")]; + fp32 var_4332_promoted = const()[name = string("op_4332_promoted"), val = fp32(0x1p+1)]; + tensor var_4338 = pow(x = hidden_states_223, y = var_4332_promoted)[name = string("op_4338")]; + tensor variance_75_axes_0 = const()[name = string("variance_75_axes_0"), val = tensor([-1])]; + bool variance_75_keep_dims_0 = const()[name = string("variance_75_keep_dims_0"), val = bool(true)]; + tensor variance_75 = reduce_mean(axes = variance_75_axes_0, keep_dims = variance_75_keep_dims_0, x = var_4338)[name = string("variance_75")]; + fp32 var_4341 = const()[name = string("op_4341"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4342 = add(x = variance_75, y = var_4341)[name = string("op_4342")]; + fp32 var_4343_epsilon_0 = const()[name = string("op_4343_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4343 = rsqrt(epsilon = var_4343_epsilon_0, x = var_4342)[name = string("op_4343")]; + tensor hidden_states_227 = mul(x = hidden_states_223, y = var_4343)[name = string("hidden_states_227")]; + tensor const_92 = const()[name = string("const_92"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774317760)))]; + tensor q_75 = mul(x = const_92, y = hidden_states_227)[name = string("q_75")]; + fp32 var_4350_promoted = const()[name = string("op_4350_promoted"), val = fp32(0x1p+1)]; + tensor var_4356 = pow(x = hidden_states_229, y = var_4350_promoted)[name = string("op_4356")]; + tensor variance_77_axes_0 = const()[name = string("variance_77_axes_0"), val = tensor([-1])]; + bool variance_77_keep_dims_0 = const()[name = string("variance_77_keep_dims_0"), val = bool(true)]; + tensor variance_77 = reduce_mean(axes = variance_77_axes_0, keep_dims = variance_77_keep_dims_0, x = var_4356)[name = string("variance_77")]; + fp32 var_4359 = const()[name = string("op_4359"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4360 = add(x = variance_77, y = var_4359)[name = string("op_4360")]; + fp32 var_4361_epsilon_0 = const()[name = string("op_4361_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4361 = rsqrt(epsilon = var_4361_epsilon_0, x = var_4360)[name = string("op_4361")]; + tensor hidden_states_233 = mul(x = hidden_states_229, y = var_4361)[name = string("hidden_states_233")]; + tensor const_93 = const()[name = string("const_93"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774318336)))]; + tensor k_75 = mul(x = const_93, y = hidden_states_233)[name = string("k_75")]; + tensor q_77_perm_0 = const()[name = string("q_77_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_77_perm_0 = const()[name = string("k_77_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_77 = transpose(perm = q_77_perm_0, x = q_75)[name = string("transpose_75")]; + tensor var_4378 = mul(x = q_77, y = cos_3)[name = string("op_4378")]; + tensor x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_37 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_77)[name = string("x1_37")]; + tensor x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_37 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_77)[name = string("x2_37")]; + fp32 const_96_promoted = const()[name = string("const_96_promoted"), val = fp32(-0x1p+0)]; + tensor var_4399 = mul(x = x2_37, y = const_96_promoted)[name = string("op_4399")]; + int32 var_4401 = const()[name = string("op_4401"), val = int32(-1)]; + bool var_4402_interleave_0 = const()[name = string("op_4402_interleave_0"), val = bool(false)]; + tensor var_4402 = concat(axis = var_4401, interleave = var_4402_interleave_0, values = (var_4399, x1_37))[name = string("op_4402")]; + tensor var_4403 = mul(x = var_4402, y = sin_3)[name = string("op_4403")]; + tensor q_79 = add(x = var_4378, y = var_4403)[name = string("q_79")]; + tensor k_77 = transpose(perm = k_77_perm_0, x = k_75)[name = string("transpose_74")]; + tensor var_4406 = mul(x = k_77, y = cos_3)[name = string("op_4406")]; + tensor x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_39 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_77)[name = string("x1_39")]; + tensor x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_39 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_77)[name = string("x2_39")]; + fp32 const_99_promoted = const()[name = string("const_99_promoted"), val = fp32(-0x1p+0)]; + tensor var_4427 = mul(x = x2_39, y = const_99_promoted)[name = string("op_4427")]; + int32 var_4429 = const()[name = string("op_4429"), val = int32(-1)]; + bool var_4430_interleave_0 = const()[name = string("op_4430_interleave_0"), val = bool(false)]; + tensor var_4430 = concat(axis = var_4429, interleave = var_4430_interleave_0, values = (var_4427, x1_39))[name = string("op_4430")]; + tensor var_4431 = mul(x = var_4430, y = sin_3)[name = string("op_4431")]; + tensor k_79 = add(x = var_4406, y = var_4431)[name = string("k_79")]; + tensor var_4438 = const()[name = string("op_4438"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_19 = reshape(shape = var_4438, x = k_79)[name = string("nk_flat_19")]; + tensor var_4444 = const()[name = string("op_4444"), val = tensor([1, 1024, 1, 1])]; + tensor v_59 = transpose(perm = v_59_perm_0, x = v_57)[name = string("transpose_73")]; + tensor nv_flat_19 = reshape(shape = var_4444, x = v_59)[name = string("nv_flat_19")]; + tensor var_4453 = mul(x = var_4247, y = var_1194)[name = string("op_4453")]; + tensor var_4454 = mul(x = nk_flat_19, y = update_mask_1)[name = string("op_4454")]; + tensor key_cache_41 = add(x = var_4453, y = var_4454)[name = string("key_cache_41")]; + tensor var_4460 = mul(x = var_4267, y = var_1194)[name = string("op_4460")]; + tensor var_4461 = mul(x = nv_flat_19, y = update_mask_1)[name = string("op_4461")]; + tensor value_cache_41 = add(x = var_4460, y = var_4461)[name = string("value_cache_41")]; + tensor kc_55_axes_0 = const()[name = string("kc_55_axes_0"), val = tensor([2])]; + tensor kc_55 = squeeze(axes = kc_55_axes_0, x = key_cache_41)[name = string("kc_55")]; + tensor var_4470 = const()[name = string("op_4470"), val = tensor([1, 8, 128, 256])]; + tensor kc_57 = reshape(shape = var_4470, x = kc_55)[name = string("kc_57")]; + tensor vc_55_axes_0 = const()[name = string("vc_55_axes_0"), val = tensor([2])]; + tensor vc_55 = squeeze(axes = vc_55_axes_0, x = value_cache_41)[name = string("vc_55")]; + tensor var_4478 = const()[name = string("op_4478"), val = tensor([1, 8, 128, 256])]; + tensor vc_57 = reshape(shape = var_4478, x = vc_55)[name = string("vc_57")]; + tensor var_4481_axes_0 = const()[name = string("op_4481_axes_0"), val = tensor([2])]; + tensor var_4481 = expand_dims(axes = var_4481_axes_0, x = kc_57)[name = string("op_4481")]; + tensor var_4489_reps_0 = const()[name = string("op_4489_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_4489 = tile(reps = var_4489_reps_0, x = var_4481)[name = string("op_4489")]; + tensor var_4494 = const()[name = string("op_4494"), val = tensor([1, 16, 128, 256])]; + tensor kc_59 = reshape(shape = var_4494, x = var_4489)[name = string("kc_59")]; + tensor var_4497_axes_0 = const()[name = string("op_4497_axes_0"), val = tensor([2])]; + tensor var_4497 = expand_dims(axes = var_4497_axes_0, x = vc_57)[name = string("op_4497")]; + tensor var_4505_reps_0 = const()[name = string("op_4505_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_4505 = tile(reps = var_4505_reps_0, x = var_4497)[name = string("op_4505")]; + tensor var_4510 = const()[name = string("op_4510"), val = tensor([1, 16, 128, 256])]; + tensor vc_59 = reshape(shape = var_4510, x = var_4505)[name = string("vc_59")]; + bool var_4512_transpose_x_0 = const()[name = string("op_4512_transpose_x_0"), val = bool(false)]; + bool var_4512_transpose_y_0 = const()[name = string("op_4512_transpose_y_0"), val = bool(false)]; + tensor var_4512 = matmul(transpose_x = var_4512_transpose_x_0, transpose_y = var_4512_transpose_y_0, x = q_79, y = kc_59)[name = string("op_4512")]; + fp32 _inversed_attn_weights_73_y_0 = const()[name = string("_inversed_attn_weights_73_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_73 = mul(x = var_4512, y = _inversed_attn_weights_73_y_0)[name = string("_inversed_attn_weights_73")]; + tensor attn_weights_75 = add(x = _inversed_attn_weights_73, y = mask_1)[name = string("attn_weights_75")]; + int32 var_4526 = const()[name = string("op_4526"), val = int32(-1)]; + tensor attn_weights_79 = softmax(axis = var_4526, x = attn_weights_75)[name = string("attn_weights_79")]; + bool attn_output_37_transpose_x_1 = const()[name = string("attn_output_37_transpose_x_1"), val = bool(false)]; + bool attn_output_37_transpose_y_1 = const()[name = string("attn_output_37_transpose_y_1"), val = bool(true)]; + tensor attn_output_37 = matmul(transpose_x = attn_output_37_transpose_x_1, transpose_y = attn_output_37_transpose_y_1, x = attn_weights_79, y = vc_59)[name = string("attn_output_37")]; + tensor var_4535_perm_0 = const()[name = string("op_4535_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4539 = const()[name = string("op_4539"), val = tensor([1, 1, -1])]; + tensor var_4535 = transpose(perm = var_4535_perm_0, x = attn_output_37)[name = string("transpose_72")]; + tensor input_93 = reshape(shape = var_4539, x = var_4535)[name = string("input_93")]; + tensor attn_output_39 = linear(bias = linear_1_bias_0, weight = layers_9_self_attn_o_proj_weight, x = input_93)[name = string("linear_66")]; + tensor var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor([0])]; + tensor var_4545 = squeeze(axes = var_4545_axes_0, x = attn_output_39)[name = string("op_4545")]; + tensor var_4547_axes_0 = const()[name = string("op_4547_axes_0"), val = tensor([0])]; + tensor var_4547 = squeeze(axes = var_4547_axes_0, x = var_4545)[name = string("op_4547")]; + tensor var_4549_axes_0 = const()[name = string("op_4549_axes_0"), val = tensor([-1])]; + tensor var_4549 = expand_dims(axes = var_4549_axes_0, x = var_4547)[name = string("op_4549")]; + tensor attn_4d_19_axes_0 = const()[name = string("attn_4d_19_axes_0"), val = tensor([-1])]; + tensor attn_4d_19 = expand_dims(axes = attn_4d_19_axes_0, x = var_4549)[name = string("attn_4d_19")]; + tensor hidden_37 = add(x = hidden_35, y = attn_4d_19)[name = string("hidden_37")]; + tensor var_4555_axes_0 = const()[name = string("op_4555_axes_0"), val = tensor([-1])]; + tensor var_4555 = squeeze(axes = var_4555_axes_0, x = hidden_37)[name = string("op_4555")]; + tensor var_4557_axes_0 = const()[name = string("op_4557_axes_0"), val = tensor([-1])]; + tensor var_4557 = squeeze(axes = var_4557_axes_0, x = var_4555)[name = string("op_4557")]; + tensor hidden_states_235_axes_0 = const()[name = string("hidden_states_235_axes_0"), val = tensor([0])]; + tensor hidden_states_235 = expand_dims(axes = hidden_states_235_axes_0, x = var_4557)[name = string("hidden_states_235")]; + fp32 var_4563_promoted = const()[name = string("op_4563_promoted"), val = fp32(0x1p+1)]; + tensor var_4569 = pow(x = hidden_states_235, y = var_4563_promoted)[name = string("op_4569")]; + tensor variance_79_axes_0 = const()[name = string("variance_79_axes_0"), val = tensor([-1])]; + bool variance_79_keep_dims_0 = const()[name = string("variance_79_keep_dims_0"), val = bool(true)]; + tensor variance_79 = reduce_mean(axes = variance_79_axes_0, keep_dims = variance_79_keep_dims_0, x = var_4569)[name = string("variance_79")]; + fp32 var_4572 = const()[name = string("op_4572"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4573 = add(x = variance_79, y = var_4572)[name = string("op_4573")]; + fp32 var_4574_epsilon_0 = const()[name = string("op_4574_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4574 = rsqrt(epsilon = var_4574_epsilon_0, x = var_4573)[name = string("op_4574")]; + tensor hidden_states_239 = mul(x = hidden_states_235, y = var_4574)[name = string("hidden_states_239")]; + tensor const_100 = const()[name = string("const_100"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774318912)))]; + tensor input_95 = mul(x = const_100, y = hidden_states_239)[name = string("input_95")]; + tensor input_97 = linear(bias = linear_4_bias_0, weight = layers_9_mlp_gate_proj_weight, x = input_95)[name = string("linear_67")]; + tensor var_4584 = silu(x = input_97)[name = string("op_4584")]; + tensor var_4586 = linear(bias = linear_4_bias_0, weight = layers_9_mlp_up_proj_weight, x = input_95)[name = string("linear_68")]; + tensor input_99 = mul(x = var_4584, y = var_4586)[name = string("input_99")]; + tensor mlp_out_19 = linear(bias = linear_1_bias_0, weight = layers_9_mlp_down_proj_weight, x = input_99)[name = string("linear_69")]; + tensor var_4591_axes_0 = const()[name = string("op_4591_axes_0"), val = tensor([0])]; + tensor var_4591 = squeeze(axes = var_4591_axes_0, x = mlp_out_19)[name = string("op_4591")]; + tensor var_4593_axes_0 = const()[name = string("op_4593_axes_0"), val = tensor([0])]; + tensor var_4593 = squeeze(axes = var_4593_axes_0, x = var_4591)[name = string("op_4593")]; + tensor var_4595_axes_0 = const()[name = string("op_4595_axes_0"), val = tensor([-1])]; + tensor var_4595 = expand_dims(axes = var_4595_axes_0, x = var_4593)[name = string("op_4595")]; + tensor mlp_4d_19_axes_0 = const()[name = string("mlp_4d_19_axes_0"), val = tensor([-1])]; + tensor mlp_4d_19 = expand_dims(axes = mlp_4d_19_axes_0, x = var_4595)[name = string("mlp_4d_19")]; + tensor hidden_39 = add(x = hidden_37, y = mlp_4d_19)[name = string("hidden_39")]; + tensor var_4609_begin_0 = const()[name = string("op_4609_begin_0"), val = tensor([0, 10240, 0, 0])]; + tensor var_4609_end_0 = const()[name = string("op_4609_end_0"), val = tensor([1, 11264, 1, 256])]; + tensor var_4609_end_mask_0 = const()[name = string("op_4609_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4609 = slice_by_index(begin = var_4609_begin_0, end = var_4609_end_0, end_mask = var_4609_end_mask_0, x = cast_3)[name = string("op_4609")]; + tensor var_4629_begin_0 = const()[name = string("op_4629_begin_0"), val = tensor([0, 10240, 0, 0])]; + tensor var_4629_end_0 = const()[name = string("op_4629_end_0"), val = tensor([1, 11264, 1, 256])]; + tensor var_4629_end_mask_0 = const()[name = string("op_4629_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4629 = slice_by_index(begin = var_4629_begin_0, end = var_4629_end_0, end_mask = var_4629_end_mask_0, x = cast_4)[name = string("op_4629")]; + tensor var_4641_axes_0 = const()[name = string("op_4641_axes_0"), val = tensor([-1])]; + tensor var_4641 = squeeze(axes = var_4641_axes_0, x = hidden_39)[name = string("op_4641")]; + tensor var_4643_axes_0 = const()[name = string("op_4643_axes_0"), val = tensor([-1])]; + tensor var_4643 = squeeze(axes = var_4643_axes_0, x = var_4641)[name = string("op_4643")]; + tensor hidden_states_241_axes_0 = const()[name = string("hidden_states_241_axes_0"), val = tensor([0])]; + tensor hidden_states_241 = expand_dims(axes = hidden_states_241_axes_0, x = var_4643)[name = string("hidden_states_241")]; + fp32 var_4649_promoted = const()[name = string("op_4649_promoted"), val = fp32(0x1p+1)]; + tensor var_4655 = pow(x = hidden_states_241, y = var_4649_promoted)[name = string("op_4655")]; + tensor variance_81_axes_0 = const()[name = string("variance_81_axes_0"), val = tensor([-1])]; + bool variance_81_keep_dims_0 = const()[name = string("variance_81_keep_dims_0"), val = bool(true)]; + tensor variance_81 = reduce_mean(axes = variance_81_axes_0, keep_dims = variance_81_keep_dims_0, x = var_4655)[name = string("variance_81")]; + fp32 var_4658 = const()[name = string("op_4658"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4659 = add(x = variance_81, y = var_4658)[name = string("op_4659")]; + fp32 var_4660_epsilon_0 = const()[name = string("op_4660_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4660 = rsqrt(epsilon = var_4660_epsilon_0, x = var_4659)[name = string("op_4660")]; + tensor hidden_states_245 = mul(x = hidden_states_241, y = var_4660)[name = string("hidden_states_245")]; + tensor const_101 = const()[name = string("const_101"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774323072)))]; + tensor input_101 = mul(x = const_101, y = hidden_states_245)[name = string("input_101")]; + tensor q_81 = linear(bias = linear_0_bias_0, weight = layers_10_self_attn_q_proj_weight, x = input_101)[name = string("linear_70")]; + tensor k_81 = linear(bias = linear_1_bias_0, weight = layers_10_self_attn_k_proj_weight, x = input_101)[name = string("linear_71")]; + tensor v_61 = linear(bias = linear_1_bias_0, weight = layers_10_self_attn_v_proj_weight, x = input_101)[name = string("linear_72")]; + tensor var_4677 = const()[name = string("op_4677"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_247 = reshape(shape = var_4677, x = q_81)[name = string("hidden_states_247")]; + tensor var_4683 = const()[name = string("op_4683"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_253 = reshape(shape = var_4683, x = k_81)[name = string("hidden_states_253")]; + tensor var_4689 = const()[name = string("op_4689"), val = tensor([1, 1, 8, 128])]; + tensor v_63 = reshape(shape = var_4689, x = v_61)[name = string("v_63")]; + fp32 var_4694_promoted = const()[name = string("op_4694_promoted"), val = fp32(0x1p+1)]; + tensor var_4700 = pow(x = hidden_states_247, y = var_4694_promoted)[name = string("op_4700")]; + tensor variance_83_axes_0 = const()[name = string("variance_83_axes_0"), val = tensor([-1])]; + bool variance_83_keep_dims_0 = const()[name = string("variance_83_keep_dims_0"), val = bool(true)]; + tensor variance_83 = reduce_mean(axes = variance_83_axes_0, keep_dims = variance_83_keep_dims_0, x = var_4700)[name = string("variance_83")]; + fp32 var_4703 = const()[name = string("op_4703"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4704 = add(x = variance_83, y = var_4703)[name = string("op_4704")]; + fp32 var_4705_epsilon_0 = const()[name = string("op_4705_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4705 = rsqrt(epsilon = var_4705_epsilon_0, x = var_4704)[name = string("op_4705")]; + tensor hidden_states_251 = mul(x = hidden_states_247, y = var_4705)[name = string("hidden_states_251")]; + tensor const_102 = const()[name = string("const_102"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774327232)))]; + tensor q_83 = mul(x = const_102, y = hidden_states_251)[name = string("q_83")]; + fp32 var_4712_promoted = const()[name = string("op_4712_promoted"), val = fp32(0x1p+1)]; + tensor var_4718 = pow(x = hidden_states_253, y = var_4712_promoted)[name = string("op_4718")]; + tensor variance_85_axes_0 = const()[name = string("variance_85_axes_0"), val = tensor([-1])]; + bool variance_85_keep_dims_0 = const()[name = string("variance_85_keep_dims_0"), val = bool(true)]; + tensor variance_85 = reduce_mean(axes = variance_85_axes_0, keep_dims = variance_85_keep_dims_0, x = var_4718)[name = string("variance_85")]; + fp32 var_4721 = const()[name = string("op_4721"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4722 = add(x = variance_85, y = var_4721)[name = string("op_4722")]; + fp32 var_4723_epsilon_0 = const()[name = string("op_4723_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4723 = rsqrt(epsilon = var_4723_epsilon_0, x = var_4722)[name = string("op_4723")]; + tensor hidden_states_257 = mul(x = hidden_states_253, y = var_4723)[name = string("hidden_states_257")]; + tensor const_103 = const()[name = string("const_103"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774327808)))]; + tensor k_83 = mul(x = const_103, y = hidden_states_257)[name = string("k_83")]; + tensor q_85_perm_0 = const()[name = string("q_85_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_85_perm_0 = const()[name = string("k_85_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_65_perm_0 = const()[name = string("v_65_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_85 = transpose(perm = q_85_perm_0, x = q_83)[name = string("transpose_71")]; + tensor var_4740 = mul(x = q_85, y = cos_3)[name = string("op_4740")]; + tensor x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_41 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_85)[name = string("x1_41")]; + tensor x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_41 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_85)[name = string("x2_41")]; + fp32 const_106_promoted = const()[name = string("const_106_promoted"), val = fp32(-0x1p+0)]; + tensor var_4761 = mul(x = x2_41, y = const_106_promoted)[name = string("op_4761")]; + int32 var_4763 = const()[name = string("op_4763"), val = int32(-1)]; + bool var_4764_interleave_0 = const()[name = string("op_4764_interleave_0"), val = bool(false)]; + tensor var_4764 = concat(axis = var_4763, interleave = var_4764_interleave_0, values = (var_4761, x1_41))[name = string("op_4764")]; + tensor var_4765 = mul(x = var_4764, y = sin_3)[name = string("op_4765")]; + tensor q_87 = add(x = var_4740, y = var_4765)[name = string("q_87")]; + tensor k_85 = transpose(perm = k_85_perm_0, x = k_83)[name = string("transpose_70")]; + tensor var_4768 = mul(x = k_85, y = cos_3)[name = string("op_4768")]; + tensor x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_43 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_85)[name = string("x1_43")]; + tensor x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_43 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_85)[name = string("x2_43")]; + fp32 const_109_promoted = const()[name = string("const_109_promoted"), val = fp32(-0x1p+0)]; + tensor var_4789 = mul(x = x2_43, y = const_109_promoted)[name = string("op_4789")]; + int32 var_4791 = const()[name = string("op_4791"), val = int32(-1)]; + bool var_4792_interleave_0 = const()[name = string("op_4792_interleave_0"), val = bool(false)]; + tensor var_4792 = concat(axis = var_4791, interleave = var_4792_interleave_0, values = (var_4789, x1_43))[name = string("op_4792")]; + tensor var_4793 = mul(x = var_4792, y = sin_3)[name = string("op_4793")]; + tensor k_87 = add(x = var_4768, y = var_4793)[name = string("k_87")]; + tensor var_4800 = const()[name = string("op_4800"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_21 = reshape(shape = var_4800, x = k_87)[name = string("nk_flat_21")]; + tensor var_4806 = const()[name = string("op_4806"), val = tensor([1, 1024, 1, 1])]; + tensor v_65 = transpose(perm = v_65_perm_0, x = v_63)[name = string("transpose_69")]; + tensor nv_flat_21 = reshape(shape = var_4806, x = v_65)[name = string("nv_flat_21")]; + tensor var_4815 = mul(x = var_4609, y = var_1194)[name = string("op_4815")]; + tensor var_4816 = mul(x = nk_flat_21, y = update_mask_1)[name = string("op_4816")]; + tensor key_cache_45 = add(x = var_4815, y = var_4816)[name = string("key_cache_45")]; + tensor var_4822 = mul(x = var_4629, y = var_1194)[name = string("op_4822")]; + tensor var_4823 = mul(x = nv_flat_21, y = update_mask_1)[name = string("op_4823")]; + tensor value_cache_45 = add(x = var_4822, y = var_4823)[name = string("value_cache_45")]; + tensor kc_61_axes_0 = const()[name = string("kc_61_axes_0"), val = tensor([2])]; + tensor kc_61 = squeeze(axes = kc_61_axes_0, x = key_cache_45)[name = string("kc_61")]; + tensor var_4832 = const()[name = string("op_4832"), val = tensor([1, 8, 128, 256])]; + tensor kc_63 = reshape(shape = var_4832, x = kc_61)[name = string("kc_63")]; + tensor vc_61_axes_0 = const()[name = string("vc_61_axes_0"), val = tensor([2])]; + tensor vc_61 = squeeze(axes = vc_61_axes_0, x = value_cache_45)[name = string("vc_61")]; + tensor var_4840 = const()[name = string("op_4840"), val = tensor([1, 8, 128, 256])]; + tensor vc_63 = reshape(shape = var_4840, x = vc_61)[name = string("vc_63")]; + tensor var_4843_axes_0 = const()[name = string("op_4843_axes_0"), val = tensor([2])]; + tensor var_4843 = expand_dims(axes = var_4843_axes_0, x = kc_63)[name = string("op_4843")]; + tensor var_4851_reps_0 = const()[name = string("op_4851_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_4851 = tile(reps = var_4851_reps_0, x = var_4843)[name = string("op_4851")]; + tensor var_4856 = const()[name = string("op_4856"), val = tensor([1, 16, 128, 256])]; + tensor kc_65 = reshape(shape = var_4856, x = var_4851)[name = string("kc_65")]; + tensor var_4859_axes_0 = const()[name = string("op_4859_axes_0"), val = tensor([2])]; + tensor var_4859 = expand_dims(axes = var_4859_axes_0, x = vc_63)[name = string("op_4859")]; + tensor var_4867_reps_0 = const()[name = string("op_4867_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_4867 = tile(reps = var_4867_reps_0, x = var_4859)[name = string("op_4867")]; + tensor var_4872 = const()[name = string("op_4872"), val = tensor([1, 16, 128, 256])]; + tensor vc_65 = reshape(shape = var_4872, x = var_4867)[name = string("vc_65")]; + bool var_4874_transpose_x_0 = const()[name = string("op_4874_transpose_x_0"), val = bool(false)]; + bool var_4874_transpose_y_0 = const()[name = string("op_4874_transpose_y_0"), val = bool(false)]; + tensor var_4874 = matmul(transpose_x = var_4874_transpose_x_0, transpose_y = var_4874_transpose_y_0, x = q_87, y = kc_65)[name = string("op_4874")]; + fp32 _inversed_attn_weights_81_y_0 = const()[name = string("_inversed_attn_weights_81_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_81 = mul(x = var_4874, y = _inversed_attn_weights_81_y_0)[name = string("_inversed_attn_weights_81")]; + tensor attn_weights_83 = add(x = _inversed_attn_weights_81, y = mask_1)[name = string("attn_weights_83")]; + int32 var_4888 = const()[name = string("op_4888"), val = int32(-1)]; + tensor attn_weights_87 = softmax(axis = var_4888, x = attn_weights_83)[name = string("attn_weights_87")]; + bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(false)]; + bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(true)]; + tensor attn_output_41 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_87, y = vc_65)[name = string("attn_output_41")]; + tensor var_4897_perm_0 = const()[name = string("op_4897_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4901 = const()[name = string("op_4901"), val = tensor([1, 1, -1])]; + tensor var_4897 = transpose(perm = var_4897_perm_0, x = attn_output_41)[name = string("transpose_68")]; + tensor input_103 = reshape(shape = var_4901, x = var_4897)[name = string("input_103")]; + tensor attn_output_43 = linear(bias = linear_1_bias_0, weight = layers_10_self_attn_o_proj_weight, x = input_103)[name = string("linear_73")]; + tensor var_4907_axes_0 = const()[name = string("op_4907_axes_0"), val = tensor([0])]; + tensor var_4907 = squeeze(axes = var_4907_axes_0, x = attn_output_43)[name = string("op_4907")]; + tensor var_4909_axes_0 = const()[name = string("op_4909_axes_0"), val = tensor([0])]; + tensor var_4909 = squeeze(axes = var_4909_axes_0, x = var_4907)[name = string("op_4909")]; + tensor var_4911_axes_0 = const()[name = string("op_4911_axes_0"), val = tensor([-1])]; + tensor var_4911 = expand_dims(axes = var_4911_axes_0, x = var_4909)[name = string("op_4911")]; + tensor attn_4d_21_axes_0 = const()[name = string("attn_4d_21_axes_0"), val = tensor([-1])]; + tensor attn_4d_21 = expand_dims(axes = attn_4d_21_axes_0, x = var_4911)[name = string("attn_4d_21")]; + tensor hidden_41 = add(x = hidden_39, y = attn_4d_21)[name = string("hidden_41")]; + tensor var_4917_axes_0 = const()[name = string("op_4917_axes_0"), val = tensor([-1])]; + tensor var_4917 = squeeze(axes = var_4917_axes_0, x = hidden_41)[name = string("op_4917")]; + tensor var_4919_axes_0 = const()[name = string("op_4919_axes_0"), val = tensor([-1])]; + tensor var_4919 = squeeze(axes = var_4919_axes_0, x = var_4917)[name = string("op_4919")]; + tensor hidden_states_259_axes_0 = const()[name = string("hidden_states_259_axes_0"), val = tensor([0])]; + tensor hidden_states_259 = expand_dims(axes = hidden_states_259_axes_0, x = var_4919)[name = string("hidden_states_259")]; + fp32 var_4925_promoted = const()[name = string("op_4925_promoted"), val = fp32(0x1p+1)]; + tensor var_4931 = pow(x = hidden_states_259, y = var_4925_promoted)[name = string("op_4931")]; + tensor variance_87_axes_0 = const()[name = string("variance_87_axes_0"), val = tensor([-1])]; + bool variance_87_keep_dims_0 = const()[name = string("variance_87_keep_dims_0"), val = bool(true)]; + tensor variance_87 = reduce_mean(axes = variance_87_axes_0, keep_dims = variance_87_keep_dims_0, x = var_4931)[name = string("variance_87")]; + fp32 var_4934 = const()[name = string("op_4934"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_4935 = add(x = variance_87, y = var_4934)[name = string("op_4935")]; + fp32 var_4936_epsilon_0 = const()[name = string("op_4936_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_4936 = rsqrt(epsilon = var_4936_epsilon_0, x = var_4935)[name = string("op_4936")]; + tensor hidden_states_263 = mul(x = hidden_states_259, y = var_4936)[name = string("hidden_states_263")]; + tensor const_110 = const()[name = string("const_110"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774328384)))]; + tensor input_105 = mul(x = const_110, y = hidden_states_263)[name = string("input_105")]; + tensor input_107 = linear(bias = linear_4_bias_0, weight = layers_10_mlp_gate_proj_weight, x = input_105)[name = string("linear_74")]; + tensor var_4946 = silu(x = input_107)[name = string("op_4946")]; + tensor var_4948 = linear(bias = linear_4_bias_0, weight = layers_10_mlp_up_proj_weight, x = input_105)[name = string("linear_75")]; + tensor input_109 = mul(x = var_4946, y = var_4948)[name = string("input_109")]; + tensor mlp_out_21 = linear(bias = linear_1_bias_0, weight = layers_10_mlp_down_proj_weight, x = input_109)[name = string("linear_76")]; + tensor var_4953_axes_0 = const()[name = string("op_4953_axes_0"), val = tensor([0])]; + tensor var_4953 = squeeze(axes = var_4953_axes_0, x = mlp_out_21)[name = string("op_4953")]; + tensor var_4955_axes_0 = const()[name = string("op_4955_axes_0"), val = tensor([0])]; + tensor var_4955 = squeeze(axes = var_4955_axes_0, x = var_4953)[name = string("op_4955")]; + tensor var_4957_axes_0 = const()[name = string("op_4957_axes_0"), val = tensor([-1])]; + tensor var_4957 = expand_dims(axes = var_4957_axes_0, x = var_4955)[name = string("op_4957")]; + tensor mlp_4d_21_axes_0 = const()[name = string("mlp_4d_21_axes_0"), val = tensor([-1])]; + tensor mlp_4d_21 = expand_dims(axes = mlp_4d_21_axes_0, x = var_4957)[name = string("mlp_4d_21")]; + tensor hidden_43 = add(x = hidden_41, y = mlp_4d_21)[name = string("hidden_43")]; + tensor var_4971_begin_0 = const()[name = string("op_4971_begin_0"), val = tensor([0, 11264, 0, 0])]; + tensor var_4971_end_0 = const()[name = string("op_4971_end_0"), val = tensor([1, 12288, 1, 256])]; + tensor var_4971_end_mask_0 = const()[name = string("op_4971_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4971 = slice_by_index(begin = var_4971_begin_0, end = var_4971_end_0, end_mask = var_4971_end_mask_0, x = cast_3)[name = string("op_4971")]; + tensor var_4991_begin_0 = const()[name = string("op_4991_begin_0"), val = tensor([0, 11264, 0, 0])]; + tensor var_4991_end_0 = const()[name = string("op_4991_end_0"), val = tensor([1, 12288, 1, 256])]; + tensor var_4991_end_mask_0 = const()[name = string("op_4991_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_4991 = slice_by_index(begin = var_4991_begin_0, end = var_4991_end_0, end_mask = var_4991_end_mask_0, x = cast_4)[name = string("op_4991")]; + tensor var_5003_axes_0 = const()[name = string("op_5003_axes_0"), val = tensor([-1])]; + tensor var_5003 = squeeze(axes = var_5003_axes_0, x = hidden_43)[name = string("op_5003")]; + tensor var_5005_axes_0 = const()[name = string("op_5005_axes_0"), val = tensor([-1])]; + tensor var_5005 = squeeze(axes = var_5005_axes_0, x = var_5003)[name = string("op_5005")]; + tensor hidden_states_265_axes_0 = const()[name = string("hidden_states_265_axes_0"), val = tensor([0])]; + tensor hidden_states_265 = expand_dims(axes = hidden_states_265_axes_0, x = var_5005)[name = string("hidden_states_265")]; + fp32 var_5011_promoted = const()[name = string("op_5011_promoted"), val = fp32(0x1p+1)]; + tensor var_5017 = pow(x = hidden_states_265, y = var_5011_promoted)[name = string("op_5017")]; + tensor variance_89_axes_0 = const()[name = string("variance_89_axes_0"), val = tensor([-1])]; + bool variance_89_keep_dims_0 = const()[name = string("variance_89_keep_dims_0"), val = bool(true)]; + tensor variance_89 = reduce_mean(axes = variance_89_axes_0, keep_dims = variance_89_keep_dims_0, x = var_5017)[name = string("variance_89")]; + fp32 var_5020 = const()[name = string("op_5020"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5021 = add(x = variance_89, y = var_5020)[name = string("op_5021")]; + fp32 var_5022_epsilon_0 = const()[name = string("op_5022_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5022 = rsqrt(epsilon = var_5022_epsilon_0, x = var_5021)[name = string("op_5022")]; + tensor hidden_states_269 = mul(x = hidden_states_265, y = var_5022)[name = string("hidden_states_269")]; + tensor const_111 = const()[name = string("const_111"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774332544)))]; + tensor input_111 = mul(x = const_111, y = hidden_states_269)[name = string("input_111")]; + tensor q_89 = linear(bias = linear_0_bias_0, weight = layers_11_self_attn_q_proj_weight, x = input_111)[name = string("linear_77")]; + tensor k_89 = linear(bias = linear_1_bias_0, weight = layers_11_self_attn_k_proj_weight, x = input_111)[name = string("linear_78")]; + tensor v_67 = linear(bias = linear_1_bias_0, weight = layers_11_self_attn_v_proj_weight, x = input_111)[name = string("linear_79")]; + tensor var_5039 = const()[name = string("op_5039"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_271 = reshape(shape = var_5039, x = q_89)[name = string("hidden_states_271")]; + tensor var_5045 = const()[name = string("op_5045"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_277 = reshape(shape = var_5045, x = k_89)[name = string("hidden_states_277")]; + tensor var_5051 = const()[name = string("op_5051"), val = tensor([1, 1, 8, 128])]; + tensor v_69 = reshape(shape = var_5051, x = v_67)[name = string("v_69")]; + fp32 var_5056_promoted = const()[name = string("op_5056_promoted"), val = fp32(0x1p+1)]; + tensor var_5062 = pow(x = hidden_states_271, y = var_5056_promoted)[name = string("op_5062")]; + tensor variance_91_axes_0 = const()[name = string("variance_91_axes_0"), val = tensor([-1])]; + bool variance_91_keep_dims_0 = const()[name = string("variance_91_keep_dims_0"), val = bool(true)]; + tensor variance_91 = reduce_mean(axes = variance_91_axes_0, keep_dims = variance_91_keep_dims_0, x = var_5062)[name = string("variance_91")]; + fp32 var_5065 = const()[name = string("op_5065"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5066 = add(x = variance_91, y = var_5065)[name = string("op_5066")]; + fp32 var_5067_epsilon_0 = const()[name = string("op_5067_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5067 = rsqrt(epsilon = var_5067_epsilon_0, x = var_5066)[name = string("op_5067")]; + tensor hidden_states_275 = mul(x = hidden_states_271, y = var_5067)[name = string("hidden_states_275")]; + tensor const_112 = const()[name = string("const_112"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774336704)))]; + tensor q_91 = mul(x = const_112, y = hidden_states_275)[name = string("q_91")]; + fp32 var_5074_promoted = const()[name = string("op_5074_promoted"), val = fp32(0x1p+1)]; + tensor var_5080 = pow(x = hidden_states_277, y = var_5074_promoted)[name = string("op_5080")]; + tensor variance_93_axes_0 = const()[name = string("variance_93_axes_0"), val = tensor([-1])]; + bool variance_93_keep_dims_0 = const()[name = string("variance_93_keep_dims_0"), val = bool(true)]; + tensor variance_93 = reduce_mean(axes = variance_93_axes_0, keep_dims = variance_93_keep_dims_0, x = var_5080)[name = string("variance_93")]; + fp32 var_5083 = const()[name = string("op_5083"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5084 = add(x = variance_93, y = var_5083)[name = string("op_5084")]; + fp32 var_5085_epsilon_0 = const()[name = string("op_5085_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5085 = rsqrt(epsilon = var_5085_epsilon_0, x = var_5084)[name = string("op_5085")]; + tensor hidden_states_281 = mul(x = hidden_states_277, y = var_5085)[name = string("hidden_states_281")]; + tensor const_113 = const()[name = string("const_113"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774337280)))]; + tensor k_91 = mul(x = const_113, y = hidden_states_281)[name = string("k_91")]; + tensor q_93_perm_0 = const()[name = string("q_93_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_93_perm_0 = const()[name = string("k_93_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_93 = transpose(perm = q_93_perm_0, x = q_91)[name = string("transpose_67")]; + tensor var_5102 = mul(x = q_93, y = cos_3)[name = string("op_5102")]; + tensor x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_45 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_93)[name = string("x1_45")]; + tensor x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_45 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_93)[name = string("x2_45")]; + fp32 const_116_promoted = const()[name = string("const_116_promoted"), val = fp32(-0x1p+0)]; + tensor var_5123 = mul(x = x2_45, y = const_116_promoted)[name = string("op_5123")]; + int32 var_5125 = const()[name = string("op_5125"), val = int32(-1)]; + bool var_5126_interleave_0 = const()[name = string("op_5126_interleave_0"), val = bool(false)]; + tensor var_5126 = concat(axis = var_5125, interleave = var_5126_interleave_0, values = (var_5123, x1_45))[name = string("op_5126")]; + tensor var_5127 = mul(x = var_5126, y = sin_3)[name = string("op_5127")]; + tensor q_95 = add(x = var_5102, y = var_5127)[name = string("q_95")]; + tensor k_93 = transpose(perm = k_93_perm_0, x = k_91)[name = string("transpose_66")]; + tensor var_5130 = mul(x = k_93, y = cos_3)[name = string("op_5130")]; + tensor x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_47 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_93)[name = string("x1_47")]; + tensor x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_47 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_93)[name = string("x2_47")]; + fp32 const_119_promoted = const()[name = string("const_119_promoted"), val = fp32(-0x1p+0)]; + tensor var_5151 = mul(x = x2_47, y = const_119_promoted)[name = string("op_5151")]; + int32 var_5153 = const()[name = string("op_5153"), val = int32(-1)]; + bool var_5154_interleave_0 = const()[name = string("op_5154_interleave_0"), val = bool(false)]; + tensor var_5154 = concat(axis = var_5153, interleave = var_5154_interleave_0, values = (var_5151, x1_47))[name = string("op_5154")]; + tensor var_5155 = mul(x = var_5154, y = sin_3)[name = string("op_5155")]; + tensor k_95 = add(x = var_5130, y = var_5155)[name = string("k_95")]; + tensor var_5162 = const()[name = string("op_5162"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_23 = reshape(shape = var_5162, x = k_95)[name = string("nk_flat_23")]; + tensor var_5168 = const()[name = string("op_5168"), val = tensor([1, 1024, 1, 1])]; + tensor v_71 = transpose(perm = v_71_perm_0, x = v_69)[name = string("transpose_65")]; + tensor nv_flat_23 = reshape(shape = var_5168, x = v_71)[name = string("nv_flat_23")]; + tensor var_5177 = mul(x = var_4971, y = var_1194)[name = string("op_5177")]; + tensor var_5178 = mul(x = nk_flat_23, y = update_mask_1)[name = string("op_5178")]; + tensor key_cache_49 = add(x = var_5177, y = var_5178)[name = string("key_cache_49")]; + tensor var_5184 = mul(x = var_4991, y = var_1194)[name = string("op_5184")]; + tensor var_5185 = mul(x = nv_flat_23, y = update_mask_1)[name = string("op_5185")]; + tensor value_cache_49 = add(x = var_5184, y = var_5185)[name = string("value_cache_49")]; + tensor kc_67_axes_0 = const()[name = string("kc_67_axes_0"), val = tensor([2])]; + tensor kc_67 = squeeze(axes = kc_67_axes_0, x = key_cache_49)[name = string("kc_67")]; + tensor var_5194 = const()[name = string("op_5194"), val = tensor([1, 8, 128, 256])]; + tensor kc_69 = reshape(shape = var_5194, x = kc_67)[name = string("kc_69")]; + tensor vc_67_axes_0 = const()[name = string("vc_67_axes_0"), val = tensor([2])]; + tensor vc_67 = squeeze(axes = vc_67_axes_0, x = value_cache_49)[name = string("vc_67")]; + tensor var_5202 = const()[name = string("op_5202"), val = tensor([1, 8, 128, 256])]; + tensor vc_69 = reshape(shape = var_5202, x = vc_67)[name = string("vc_69")]; + tensor var_5205_axes_0 = const()[name = string("op_5205_axes_0"), val = tensor([2])]; + tensor var_5205 = expand_dims(axes = var_5205_axes_0, x = kc_69)[name = string("op_5205")]; + tensor var_5213_reps_0 = const()[name = string("op_5213_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_5213 = tile(reps = var_5213_reps_0, x = var_5205)[name = string("op_5213")]; + tensor var_5218 = const()[name = string("op_5218"), val = tensor([1, 16, 128, 256])]; + tensor kc_71 = reshape(shape = var_5218, x = var_5213)[name = string("kc_71")]; + tensor var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor([2])]; + tensor var_5221 = expand_dims(axes = var_5221_axes_0, x = vc_69)[name = string("op_5221")]; + tensor var_5229_reps_0 = const()[name = string("op_5229_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_5229 = tile(reps = var_5229_reps_0, x = var_5221)[name = string("op_5229")]; + tensor var_5234 = const()[name = string("op_5234"), val = tensor([1, 16, 128, 256])]; + tensor vc_71 = reshape(shape = var_5234, x = var_5229)[name = string("vc_71")]; + bool var_5236_transpose_x_0 = const()[name = string("op_5236_transpose_x_0"), val = bool(false)]; + bool var_5236_transpose_y_0 = const()[name = string("op_5236_transpose_y_0"), val = bool(false)]; + tensor var_5236 = matmul(transpose_x = var_5236_transpose_x_0, transpose_y = var_5236_transpose_y_0, x = q_95, y = kc_71)[name = string("op_5236")]; + fp32 _inversed_attn_weights_89_y_0 = const()[name = string("_inversed_attn_weights_89_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_89 = mul(x = var_5236, y = _inversed_attn_weights_89_y_0)[name = string("_inversed_attn_weights_89")]; + tensor attn_weights_91 = add(x = _inversed_attn_weights_89, y = mask_1)[name = string("attn_weights_91")]; + int32 var_5250 = const()[name = string("op_5250"), val = int32(-1)]; + tensor attn_weights_95 = softmax(axis = var_5250, x = attn_weights_91)[name = string("attn_weights_95")]; + bool attn_output_45_transpose_x_1 = const()[name = string("attn_output_45_transpose_x_1"), val = bool(false)]; + bool attn_output_45_transpose_y_1 = const()[name = string("attn_output_45_transpose_y_1"), val = bool(true)]; + tensor attn_output_45 = matmul(transpose_x = attn_output_45_transpose_x_1, transpose_y = attn_output_45_transpose_y_1, x = attn_weights_95, y = vc_71)[name = string("attn_output_45")]; + tensor var_5259_perm_0 = const()[name = string("op_5259_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5263 = const()[name = string("op_5263"), val = tensor([1, 1, -1])]; + tensor var_5259 = transpose(perm = var_5259_perm_0, x = attn_output_45)[name = string("transpose_64")]; + tensor input_113 = reshape(shape = var_5263, x = var_5259)[name = string("input_113")]; + tensor attn_output_47 = linear(bias = linear_1_bias_0, weight = layers_11_self_attn_o_proj_weight, x = input_113)[name = string("linear_80")]; + tensor var_5269_axes_0 = const()[name = string("op_5269_axes_0"), val = tensor([0])]; + tensor var_5269 = squeeze(axes = var_5269_axes_0, x = attn_output_47)[name = string("op_5269")]; + tensor var_5271_axes_0 = const()[name = string("op_5271_axes_0"), val = tensor([0])]; + tensor var_5271 = squeeze(axes = var_5271_axes_0, x = var_5269)[name = string("op_5271")]; + tensor var_5273_axes_0 = const()[name = string("op_5273_axes_0"), val = tensor([-1])]; + tensor var_5273 = expand_dims(axes = var_5273_axes_0, x = var_5271)[name = string("op_5273")]; + tensor attn_4d_23_axes_0 = const()[name = string("attn_4d_23_axes_0"), val = tensor([-1])]; + tensor attn_4d_23 = expand_dims(axes = attn_4d_23_axes_0, x = var_5273)[name = string("attn_4d_23")]; + tensor hidden_45 = add(x = hidden_43, y = attn_4d_23)[name = string("hidden_45")]; + tensor var_5279_axes_0 = const()[name = string("op_5279_axes_0"), val = tensor([-1])]; + tensor var_5279 = squeeze(axes = var_5279_axes_0, x = hidden_45)[name = string("op_5279")]; + tensor var_5281_axes_0 = const()[name = string("op_5281_axes_0"), val = tensor([-1])]; + tensor var_5281 = squeeze(axes = var_5281_axes_0, x = var_5279)[name = string("op_5281")]; + tensor hidden_states_283_axes_0 = const()[name = string("hidden_states_283_axes_0"), val = tensor([0])]; + tensor hidden_states_283 = expand_dims(axes = hidden_states_283_axes_0, x = var_5281)[name = string("hidden_states_283")]; + fp32 var_5287_promoted = const()[name = string("op_5287_promoted"), val = fp32(0x1p+1)]; + tensor var_5293 = pow(x = hidden_states_283, y = var_5287_promoted)[name = string("op_5293")]; + tensor variance_95_axes_0 = const()[name = string("variance_95_axes_0"), val = tensor([-1])]; + bool variance_95_keep_dims_0 = const()[name = string("variance_95_keep_dims_0"), val = bool(true)]; + tensor variance_95 = reduce_mean(axes = variance_95_axes_0, keep_dims = variance_95_keep_dims_0, x = var_5293)[name = string("variance_95")]; + fp32 var_5296 = const()[name = string("op_5296"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5297 = add(x = variance_95, y = var_5296)[name = string("op_5297")]; + fp32 var_5298_epsilon_0 = const()[name = string("op_5298_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5298 = rsqrt(epsilon = var_5298_epsilon_0, x = var_5297)[name = string("op_5298")]; + tensor hidden_states_287 = mul(x = hidden_states_283, y = var_5298)[name = string("hidden_states_287")]; + tensor const_120 = const()[name = string("const_120"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774337856)))]; + tensor input_115 = mul(x = const_120, y = hidden_states_287)[name = string("input_115")]; + tensor input_117 = linear(bias = linear_4_bias_0, weight = layers_11_mlp_gate_proj_weight, x = input_115)[name = string("linear_81")]; + tensor var_5308 = silu(x = input_117)[name = string("op_5308")]; + tensor var_5310 = linear(bias = linear_4_bias_0, weight = layers_11_mlp_up_proj_weight, x = input_115)[name = string("linear_82")]; + tensor input_119 = mul(x = var_5308, y = var_5310)[name = string("input_119")]; + tensor mlp_out_23 = linear(bias = linear_1_bias_0, weight = layers_11_mlp_down_proj_weight, x = input_119)[name = string("linear_83")]; + tensor var_5315_axes_0 = const()[name = string("op_5315_axes_0"), val = tensor([0])]; + tensor var_5315 = squeeze(axes = var_5315_axes_0, x = mlp_out_23)[name = string("op_5315")]; + tensor var_5317_axes_0 = const()[name = string("op_5317_axes_0"), val = tensor([0])]; + tensor var_5317 = squeeze(axes = var_5317_axes_0, x = var_5315)[name = string("op_5317")]; + tensor var_5319_axes_0 = const()[name = string("op_5319_axes_0"), val = tensor([-1])]; + tensor var_5319 = expand_dims(axes = var_5319_axes_0, x = var_5317)[name = string("op_5319")]; + tensor mlp_4d_23_axes_0 = const()[name = string("mlp_4d_23_axes_0"), val = tensor([-1])]; + tensor mlp_4d_23 = expand_dims(axes = mlp_4d_23_axes_0, x = var_5319)[name = string("mlp_4d_23")]; + tensor hidden_47 = add(x = hidden_45, y = mlp_4d_23)[name = string("hidden_47")]; + tensor var_5333_begin_0 = const()[name = string("op_5333_begin_0"), val = tensor([0, 12288, 0, 0])]; + tensor var_5333_end_0 = const()[name = string("op_5333_end_0"), val = tensor([1, 13312, 1, 256])]; + tensor var_5333_end_mask_0 = const()[name = string("op_5333_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5333 = slice_by_index(begin = var_5333_begin_0, end = var_5333_end_0, end_mask = var_5333_end_mask_0, x = cast_3)[name = string("op_5333")]; + tensor var_5353_begin_0 = const()[name = string("op_5353_begin_0"), val = tensor([0, 12288, 0, 0])]; + tensor var_5353_end_0 = const()[name = string("op_5353_end_0"), val = tensor([1, 13312, 1, 256])]; + tensor var_5353_end_mask_0 = const()[name = string("op_5353_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5353 = slice_by_index(begin = var_5353_begin_0, end = var_5353_end_0, end_mask = var_5353_end_mask_0, x = cast_4)[name = string("op_5353")]; + tensor var_5365_axes_0 = const()[name = string("op_5365_axes_0"), val = tensor([-1])]; + tensor var_5365 = squeeze(axes = var_5365_axes_0, x = hidden_47)[name = string("op_5365")]; + tensor var_5367_axes_0 = const()[name = string("op_5367_axes_0"), val = tensor([-1])]; + tensor var_5367 = squeeze(axes = var_5367_axes_0, x = var_5365)[name = string("op_5367")]; + tensor hidden_states_289_axes_0 = const()[name = string("hidden_states_289_axes_0"), val = tensor([0])]; + tensor hidden_states_289 = expand_dims(axes = hidden_states_289_axes_0, x = var_5367)[name = string("hidden_states_289")]; + fp32 var_5373_promoted = const()[name = string("op_5373_promoted"), val = fp32(0x1p+1)]; + tensor var_5379 = pow(x = hidden_states_289, y = var_5373_promoted)[name = string("op_5379")]; + tensor variance_97_axes_0 = const()[name = string("variance_97_axes_0"), val = tensor([-1])]; + bool variance_97_keep_dims_0 = const()[name = string("variance_97_keep_dims_0"), val = bool(true)]; + tensor variance_97 = reduce_mean(axes = variance_97_axes_0, keep_dims = variance_97_keep_dims_0, x = var_5379)[name = string("variance_97")]; + fp32 var_5382 = const()[name = string("op_5382"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5383 = add(x = variance_97, y = var_5382)[name = string("op_5383")]; + fp32 var_5384_epsilon_0 = const()[name = string("op_5384_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5384 = rsqrt(epsilon = var_5384_epsilon_0, x = var_5383)[name = string("op_5384")]; + tensor hidden_states_293 = mul(x = hidden_states_289, y = var_5384)[name = string("hidden_states_293")]; + tensor const_121 = const()[name = string("const_121"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774342016)))]; + tensor input_121 = mul(x = const_121, y = hidden_states_293)[name = string("input_121")]; + tensor q_97 = linear(bias = linear_0_bias_0, weight = layers_12_self_attn_q_proj_weight, x = input_121)[name = string("linear_84")]; + tensor k_97 = linear(bias = linear_1_bias_0, weight = layers_12_self_attn_k_proj_weight, x = input_121)[name = string("linear_85")]; + tensor v_73 = linear(bias = linear_1_bias_0, weight = layers_12_self_attn_v_proj_weight, x = input_121)[name = string("linear_86")]; + tensor var_5401 = const()[name = string("op_5401"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_295 = reshape(shape = var_5401, x = q_97)[name = string("hidden_states_295")]; + tensor var_5407 = const()[name = string("op_5407"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_301 = reshape(shape = var_5407, x = k_97)[name = string("hidden_states_301")]; + tensor var_5413 = const()[name = string("op_5413"), val = tensor([1, 1, 8, 128])]; + tensor v_75 = reshape(shape = var_5413, x = v_73)[name = string("v_75")]; + fp32 var_5418_promoted = const()[name = string("op_5418_promoted"), val = fp32(0x1p+1)]; + tensor var_5424 = pow(x = hidden_states_295, y = var_5418_promoted)[name = string("op_5424")]; + tensor variance_99_axes_0 = const()[name = string("variance_99_axes_0"), val = tensor([-1])]; + bool variance_99_keep_dims_0 = const()[name = string("variance_99_keep_dims_0"), val = bool(true)]; + tensor variance_99 = reduce_mean(axes = variance_99_axes_0, keep_dims = variance_99_keep_dims_0, x = var_5424)[name = string("variance_99")]; + fp32 var_5427 = const()[name = string("op_5427"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5428 = add(x = variance_99, y = var_5427)[name = string("op_5428")]; + fp32 var_5429_epsilon_0 = const()[name = string("op_5429_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5429 = rsqrt(epsilon = var_5429_epsilon_0, x = var_5428)[name = string("op_5429")]; + tensor hidden_states_299 = mul(x = hidden_states_295, y = var_5429)[name = string("hidden_states_299")]; + tensor const_122 = const()[name = string("const_122"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774346176)))]; + tensor q_99 = mul(x = const_122, y = hidden_states_299)[name = string("q_99")]; + fp32 var_5436_promoted = const()[name = string("op_5436_promoted"), val = fp32(0x1p+1)]; + tensor var_5442 = pow(x = hidden_states_301, y = var_5436_promoted)[name = string("op_5442")]; + tensor variance_101_axes_0 = const()[name = string("variance_101_axes_0"), val = tensor([-1])]; + bool variance_101_keep_dims_0 = const()[name = string("variance_101_keep_dims_0"), val = bool(true)]; + tensor variance_101 = reduce_mean(axes = variance_101_axes_0, keep_dims = variance_101_keep_dims_0, x = var_5442)[name = string("variance_101")]; + fp32 var_5445 = const()[name = string("op_5445"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5446 = add(x = variance_101, y = var_5445)[name = string("op_5446")]; + fp32 var_5447_epsilon_0 = const()[name = string("op_5447_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5447 = rsqrt(epsilon = var_5447_epsilon_0, x = var_5446)[name = string("op_5447")]; + tensor hidden_states_305 = mul(x = hidden_states_301, y = var_5447)[name = string("hidden_states_305")]; + tensor const_123 = const()[name = string("const_123"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774346752)))]; + tensor k_99 = mul(x = const_123, y = hidden_states_305)[name = string("k_99")]; + tensor q_101_perm_0 = const()[name = string("q_101_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_101_perm_0 = const()[name = string("k_101_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_77_perm_0 = const()[name = string("v_77_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_101 = transpose(perm = q_101_perm_0, x = q_99)[name = string("transpose_63")]; + tensor var_5464 = mul(x = q_101, y = cos_3)[name = string("op_5464")]; + tensor x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_49 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_101)[name = string("x1_49")]; + tensor x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_49 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_101)[name = string("x2_49")]; + fp32 const_126_promoted = const()[name = string("const_126_promoted"), val = fp32(-0x1p+0)]; + tensor var_5485 = mul(x = x2_49, y = const_126_promoted)[name = string("op_5485")]; + int32 var_5487 = const()[name = string("op_5487"), val = int32(-1)]; + bool var_5488_interleave_0 = const()[name = string("op_5488_interleave_0"), val = bool(false)]; + tensor var_5488 = concat(axis = var_5487, interleave = var_5488_interleave_0, values = (var_5485, x1_49))[name = string("op_5488")]; + tensor var_5489 = mul(x = var_5488, y = sin_3)[name = string("op_5489")]; + tensor q_103 = add(x = var_5464, y = var_5489)[name = string("q_103")]; + tensor k_101 = transpose(perm = k_101_perm_0, x = k_99)[name = string("transpose_62")]; + tensor var_5492 = mul(x = k_101, y = cos_3)[name = string("op_5492")]; + tensor x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_51 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_101)[name = string("x1_51")]; + tensor x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_51 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_101)[name = string("x2_51")]; + fp32 const_129_promoted = const()[name = string("const_129_promoted"), val = fp32(-0x1p+0)]; + tensor var_5513 = mul(x = x2_51, y = const_129_promoted)[name = string("op_5513")]; + int32 var_5515 = const()[name = string("op_5515"), val = int32(-1)]; + bool var_5516_interleave_0 = const()[name = string("op_5516_interleave_0"), val = bool(false)]; + tensor var_5516 = concat(axis = var_5515, interleave = var_5516_interleave_0, values = (var_5513, x1_51))[name = string("op_5516")]; + tensor var_5517 = mul(x = var_5516, y = sin_3)[name = string("op_5517")]; + tensor k_103 = add(x = var_5492, y = var_5517)[name = string("k_103")]; + tensor var_5524 = const()[name = string("op_5524"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_25 = reshape(shape = var_5524, x = k_103)[name = string("nk_flat_25")]; + tensor var_5530 = const()[name = string("op_5530"), val = tensor([1, 1024, 1, 1])]; + tensor v_77 = transpose(perm = v_77_perm_0, x = v_75)[name = string("transpose_61")]; + tensor nv_flat_25 = reshape(shape = var_5530, x = v_77)[name = string("nv_flat_25")]; + tensor var_5539 = mul(x = var_5333, y = var_1194)[name = string("op_5539")]; + tensor var_5540 = mul(x = nk_flat_25, y = update_mask_1)[name = string("op_5540")]; + tensor key_cache_53 = add(x = var_5539, y = var_5540)[name = string("key_cache_53")]; + tensor var_5546 = mul(x = var_5353, y = var_1194)[name = string("op_5546")]; + tensor var_5547 = mul(x = nv_flat_25, y = update_mask_1)[name = string("op_5547")]; + tensor value_cache_53 = add(x = var_5546, y = var_5547)[name = string("value_cache_53")]; + tensor kc_73_axes_0 = const()[name = string("kc_73_axes_0"), val = tensor([2])]; + tensor kc_73 = squeeze(axes = kc_73_axes_0, x = key_cache_53)[name = string("kc_73")]; + tensor var_5556 = const()[name = string("op_5556"), val = tensor([1, 8, 128, 256])]; + tensor kc_75 = reshape(shape = var_5556, x = kc_73)[name = string("kc_75")]; + tensor vc_73_axes_0 = const()[name = string("vc_73_axes_0"), val = tensor([2])]; + tensor vc_73 = squeeze(axes = vc_73_axes_0, x = value_cache_53)[name = string("vc_73")]; + tensor var_5564 = const()[name = string("op_5564"), val = tensor([1, 8, 128, 256])]; + tensor vc_75 = reshape(shape = var_5564, x = vc_73)[name = string("vc_75")]; + tensor var_5567_axes_0 = const()[name = string("op_5567_axes_0"), val = tensor([2])]; + tensor var_5567 = expand_dims(axes = var_5567_axes_0, x = kc_75)[name = string("op_5567")]; + tensor var_5575_reps_0 = const()[name = string("op_5575_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_5575 = tile(reps = var_5575_reps_0, x = var_5567)[name = string("op_5575")]; + tensor var_5580 = const()[name = string("op_5580"), val = tensor([1, 16, 128, 256])]; + tensor kc_77 = reshape(shape = var_5580, x = var_5575)[name = string("kc_77")]; + tensor var_5583_axes_0 = const()[name = string("op_5583_axes_0"), val = tensor([2])]; + tensor var_5583 = expand_dims(axes = var_5583_axes_0, x = vc_75)[name = string("op_5583")]; + tensor var_5591_reps_0 = const()[name = string("op_5591_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_5591 = tile(reps = var_5591_reps_0, x = var_5583)[name = string("op_5591")]; + tensor var_5596 = const()[name = string("op_5596"), val = tensor([1, 16, 128, 256])]; + tensor vc_77 = reshape(shape = var_5596, x = var_5591)[name = string("vc_77")]; + bool var_5598_transpose_x_0 = const()[name = string("op_5598_transpose_x_0"), val = bool(false)]; + bool var_5598_transpose_y_0 = const()[name = string("op_5598_transpose_y_0"), val = bool(false)]; + tensor var_5598 = matmul(transpose_x = var_5598_transpose_x_0, transpose_y = var_5598_transpose_y_0, x = q_103, y = kc_77)[name = string("op_5598")]; + fp32 _inversed_attn_weights_97_y_0 = const()[name = string("_inversed_attn_weights_97_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_97 = mul(x = var_5598, y = _inversed_attn_weights_97_y_0)[name = string("_inversed_attn_weights_97")]; + tensor attn_weights_99 = add(x = _inversed_attn_weights_97, y = mask_1)[name = string("attn_weights_99")]; + int32 var_5612 = const()[name = string("op_5612"), val = int32(-1)]; + tensor attn_weights_103 = softmax(axis = var_5612, x = attn_weights_99)[name = string("attn_weights_103")]; + bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(false)]; + bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(true)]; + tensor attn_output_49 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_103, y = vc_77)[name = string("attn_output_49")]; + tensor var_5621_perm_0 = const()[name = string("op_5621_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5625 = const()[name = string("op_5625"), val = tensor([1, 1, -1])]; + tensor var_5621 = transpose(perm = var_5621_perm_0, x = attn_output_49)[name = string("transpose_60")]; + tensor input_123 = reshape(shape = var_5625, x = var_5621)[name = string("input_123")]; + tensor attn_output_51 = linear(bias = linear_1_bias_0, weight = layers_12_self_attn_o_proj_weight, x = input_123)[name = string("linear_87")]; + tensor var_5631_axes_0 = const()[name = string("op_5631_axes_0"), val = tensor([0])]; + tensor var_5631 = squeeze(axes = var_5631_axes_0, x = attn_output_51)[name = string("op_5631")]; + tensor var_5633_axes_0 = const()[name = string("op_5633_axes_0"), val = tensor([0])]; + tensor var_5633 = squeeze(axes = var_5633_axes_0, x = var_5631)[name = string("op_5633")]; + tensor var_5635_axes_0 = const()[name = string("op_5635_axes_0"), val = tensor([-1])]; + tensor var_5635 = expand_dims(axes = var_5635_axes_0, x = var_5633)[name = string("op_5635")]; + tensor attn_4d_25_axes_0 = const()[name = string("attn_4d_25_axes_0"), val = tensor([-1])]; + tensor attn_4d_25 = expand_dims(axes = attn_4d_25_axes_0, x = var_5635)[name = string("attn_4d_25")]; + tensor hidden_49 = add(x = hidden_47, y = attn_4d_25)[name = string("hidden_49")]; + tensor var_5641_axes_0 = const()[name = string("op_5641_axes_0"), val = tensor([-1])]; + tensor var_5641 = squeeze(axes = var_5641_axes_0, x = hidden_49)[name = string("op_5641")]; + tensor var_5643_axes_0 = const()[name = string("op_5643_axes_0"), val = tensor([-1])]; + tensor var_5643 = squeeze(axes = var_5643_axes_0, x = var_5641)[name = string("op_5643")]; + tensor hidden_states_307_axes_0 = const()[name = string("hidden_states_307_axes_0"), val = tensor([0])]; + tensor hidden_states_307 = expand_dims(axes = hidden_states_307_axes_0, x = var_5643)[name = string("hidden_states_307")]; + fp32 var_5649_promoted = const()[name = string("op_5649_promoted"), val = fp32(0x1p+1)]; + tensor var_5655 = pow(x = hidden_states_307, y = var_5649_promoted)[name = string("op_5655")]; + tensor variance_103_axes_0 = const()[name = string("variance_103_axes_0"), val = tensor([-1])]; + bool variance_103_keep_dims_0 = const()[name = string("variance_103_keep_dims_0"), val = bool(true)]; + tensor variance_103 = reduce_mean(axes = variance_103_axes_0, keep_dims = variance_103_keep_dims_0, x = var_5655)[name = string("variance_103")]; + fp32 var_5658 = const()[name = string("op_5658"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5659 = add(x = variance_103, y = var_5658)[name = string("op_5659")]; + fp32 var_5660_epsilon_0 = const()[name = string("op_5660_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5660 = rsqrt(epsilon = var_5660_epsilon_0, x = var_5659)[name = string("op_5660")]; + tensor hidden_states_311 = mul(x = hidden_states_307, y = var_5660)[name = string("hidden_states_311")]; + tensor const_130 = const()[name = string("const_130"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774347328)))]; + tensor input_125 = mul(x = const_130, y = hidden_states_311)[name = string("input_125")]; + tensor input_127 = linear(bias = linear_4_bias_0, weight = layers_12_mlp_gate_proj_weight, x = input_125)[name = string("linear_88")]; + tensor var_5670 = silu(x = input_127)[name = string("op_5670")]; + tensor var_5672 = linear(bias = linear_4_bias_0, weight = layers_12_mlp_up_proj_weight, x = input_125)[name = string("linear_89")]; + tensor input_129 = mul(x = var_5670, y = var_5672)[name = string("input_129")]; + tensor mlp_out_25 = linear(bias = linear_1_bias_0, weight = layers_12_mlp_down_proj_weight, x = input_129)[name = string("linear_90")]; + tensor var_5677_axes_0 = const()[name = string("op_5677_axes_0"), val = tensor([0])]; + tensor var_5677 = squeeze(axes = var_5677_axes_0, x = mlp_out_25)[name = string("op_5677")]; + tensor var_5679_axes_0 = const()[name = string("op_5679_axes_0"), val = tensor([0])]; + tensor var_5679 = squeeze(axes = var_5679_axes_0, x = var_5677)[name = string("op_5679")]; + tensor var_5681_axes_0 = const()[name = string("op_5681_axes_0"), val = tensor([-1])]; + tensor var_5681 = expand_dims(axes = var_5681_axes_0, x = var_5679)[name = string("op_5681")]; + tensor mlp_4d_25_axes_0 = const()[name = string("mlp_4d_25_axes_0"), val = tensor([-1])]; + tensor mlp_4d_25 = expand_dims(axes = mlp_4d_25_axes_0, x = var_5681)[name = string("mlp_4d_25")]; + tensor hidden_51 = add(x = hidden_49, y = mlp_4d_25)[name = string("hidden_51")]; + tensor var_5695_begin_0 = const()[name = string("op_5695_begin_0"), val = tensor([0, 13312, 0, 0])]; + tensor var_5695_end_0 = const()[name = string("op_5695_end_0"), val = tensor([1, 14336, 1, 256])]; + tensor var_5695_end_mask_0 = const()[name = string("op_5695_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5695 = slice_by_index(begin = var_5695_begin_0, end = var_5695_end_0, end_mask = var_5695_end_mask_0, x = cast_3)[name = string("op_5695")]; + tensor var_5715_begin_0 = const()[name = string("op_5715_begin_0"), val = tensor([0, 13312, 0, 0])]; + tensor var_5715_end_0 = const()[name = string("op_5715_end_0"), val = tensor([1, 14336, 1, 256])]; + tensor var_5715_end_mask_0 = const()[name = string("op_5715_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5715 = slice_by_index(begin = var_5715_begin_0, end = var_5715_end_0, end_mask = var_5715_end_mask_0, x = cast_4)[name = string("op_5715")]; + tensor var_5727_axes_0 = const()[name = string("op_5727_axes_0"), val = tensor([-1])]; + tensor var_5727 = squeeze(axes = var_5727_axes_0, x = hidden_51)[name = string("op_5727")]; + tensor var_5729_axes_0 = const()[name = string("op_5729_axes_0"), val = tensor([-1])]; + tensor var_5729 = squeeze(axes = var_5729_axes_0, x = var_5727)[name = string("op_5729")]; + tensor hidden_states_313_axes_0 = const()[name = string("hidden_states_313_axes_0"), val = tensor([0])]; + tensor hidden_states_313 = expand_dims(axes = hidden_states_313_axes_0, x = var_5729)[name = string("hidden_states_313")]; + fp32 var_5735_promoted = const()[name = string("op_5735_promoted"), val = fp32(0x1p+1)]; + tensor var_5741 = pow(x = hidden_states_313, y = var_5735_promoted)[name = string("op_5741")]; + tensor variance_105_axes_0 = const()[name = string("variance_105_axes_0"), val = tensor([-1])]; + bool variance_105_keep_dims_0 = const()[name = string("variance_105_keep_dims_0"), val = bool(true)]; + tensor variance_105 = reduce_mean(axes = variance_105_axes_0, keep_dims = variance_105_keep_dims_0, x = var_5741)[name = string("variance_105")]; + fp32 var_5744 = const()[name = string("op_5744"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5745 = add(x = variance_105, y = var_5744)[name = string("op_5745")]; + fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5746 = rsqrt(epsilon = var_5746_epsilon_0, x = var_5745)[name = string("op_5746")]; + tensor hidden_states_317 = mul(x = hidden_states_313, y = var_5746)[name = string("hidden_states_317")]; + tensor const_131 = const()[name = string("const_131"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774351488)))]; + tensor input_131 = mul(x = const_131, y = hidden_states_317)[name = string("input_131")]; + tensor q_105 = linear(bias = linear_0_bias_0, weight = layers_13_self_attn_q_proj_weight, x = input_131)[name = string("linear_91")]; + tensor k_105 = linear(bias = linear_1_bias_0, weight = layers_13_self_attn_k_proj_weight, x = input_131)[name = string("linear_92")]; + tensor v_79 = linear(bias = linear_1_bias_0, weight = layers_13_self_attn_v_proj_weight, x = input_131)[name = string("linear_93")]; + tensor var_5763 = const()[name = string("op_5763"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_319 = reshape(shape = var_5763, x = q_105)[name = string("hidden_states_319")]; + tensor var_5769 = const()[name = string("op_5769"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_325 = reshape(shape = var_5769, x = k_105)[name = string("hidden_states_325")]; + tensor var_5775 = const()[name = string("op_5775"), val = tensor([1, 1, 8, 128])]; + tensor v_81 = reshape(shape = var_5775, x = v_79)[name = string("v_81")]; + fp32 var_5780_promoted = const()[name = string("op_5780_promoted"), val = fp32(0x1p+1)]; + tensor var_5786 = pow(x = hidden_states_319, y = var_5780_promoted)[name = string("op_5786")]; + tensor variance_107_axes_0 = const()[name = string("variance_107_axes_0"), val = tensor([-1])]; + bool variance_107_keep_dims_0 = const()[name = string("variance_107_keep_dims_0"), val = bool(true)]; + tensor variance_107 = reduce_mean(axes = variance_107_axes_0, keep_dims = variance_107_keep_dims_0, x = var_5786)[name = string("variance_107")]; + fp32 var_5789 = const()[name = string("op_5789"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5790 = add(x = variance_107, y = var_5789)[name = string("op_5790")]; + fp32 var_5791_epsilon_0 = const()[name = string("op_5791_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5791 = rsqrt(epsilon = var_5791_epsilon_0, x = var_5790)[name = string("op_5791")]; + tensor hidden_states_323 = mul(x = hidden_states_319, y = var_5791)[name = string("hidden_states_323")]; + tensor const_132 = const()[name = string("const_132"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774355648)))]; + tensor q_107 = mul(x = const_132, y = hidden_states_323)[name = string("q_107")]; + fp32 var_5798_promoted = const()[name = string("op_5798_promoted"), val = fp32(0x1p+1)]; + tensor var_5804 = pow(x = hidden_states_325, y = var_5798_promoted)[name = string("op_5804")]; + tensor variance_109_axes_0 = const()[name = string("variance_109_axes_0"), val = tensor([-1])]; + bool variance_109_keep_dims_0 = const()[name = string("variance_109_keep_dims_0"), val = bool(true)]; + tensor variance_109 = reduce_mean(axes = variance_109_axes_0, keep_dims = variance_109_keep_dims_0, x = var_5804)[name = string("variance_109")]; + fp32 var_5807 = const()[name = string("op_5807"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_5808 = add(x = variance_109, y = var_5807)[name = string("op_5808")]; + fp32 var_5809_epsilon_0 = const()[name = string("op_5809_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_5809 = rsqrt(epsilon = var_5809_epsilon_0, x = var_5808)[name = string("op_5809")]; + tensor hidden_states_329 = mul(x = hidden_states_325, y = var_5809)[name = string("hidden_states_329")]; + tensor const_133 = const()[name = string("const_133"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774356224)))]; + tensor k_107 = mul(x = const_133, y = hidden_states_329)[name = string("k_107")]; + tensor q_109_perm_0 = const()[name = string("q_109_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_109_perm_0 = const()[name = string("k_109_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_109 = transpose(perm = q_109_perm_0, x = q_107)[name = string("transpose_59")]; + tensor var_5826 = mul(x = q_109, y = cos_3)[name = string("op_5826")]; + tensor x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_53 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_109)[name = string("x1_53")]; + tensor x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_53 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_109)[name = string("x2_53")]; + fp32 const_136_promoted = const()[name = string("const_136_promoted"), val = fp32(-0x1p+0)]; + tensor var_5847 = mul(x = x2_53, y = const_136_promoted)[name = string("op_5847")]; + int32 var_5849 = const()[name = string("op_5849"), val = int32(-1)]; + bool var_5850_interleave_0 = const()[name = string("op_5850_interleave_0"), val = bool(false)]; + tensor var_5850 = concat(axis = var_5849, interleave = var_5850_interleave_0, values = (var_5847, x1_53))[name = string("op_5850")]; + tensor var_5851 = mul(x = var_5850, y = sin_3)[name = string("op_5851")]; + tensor q_111 = add(x = var_5826, y = var_5851)[name = string("q_111")]; + tensor k_109 = transpose(perm = k_109_perm_0, x = k_107)[name = string("transpose_58")]; + tensor var_5854 = mul(x = k_109, y = cos_3)[name = string("op_5854")]; + tensor x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_55 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_109)[name = string("x1_55")]; + tensor x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_55 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_109)[name = string("x2_55")]; + fp32 const_139_promoted = const()[name = string("const_139_promoted"), val = fp32(-0x1p+0)]; + tensor var_5875 = mul(x = x2_55, y = const_139_promoted)[name = string("op_5875")]; + int32 var_5877 = const()[name = string("op_5877"), val = int32(-1)]; + bool var_5878_interleave_0 = const()[name = string("op_5878_interleave_0"), val = bool(false)]; + tensor var_5878 = concat(axis = var_5877, interleave = var_5878_interleave_0, values = (var_5875, x1_55))[name = string("op_5878")]; + tensor var_5879 = mul(x = var_5878, y = sin_3)[name = string("op_5879")]; + tensor k_111 = add(x = var_5854, y = var_5879)[name = string("k_111")]; + tensor var_5886 = const()[name = string("op_5886"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_27 = reshape(shape = var_5886, x = k_111)[name = string("nk_flat_27")]; + tensor var_5892 = const()[name = string("op_5892"), val = tensor([1, 1024, 1, 1])]; + tensor v_83 = transpose(perm = v_83_perm_0, x = v_81)[name = string("transpose_57")]; + tensor nv_flat_27 = reshape(shape = var_5892, x = v_83)[name = string("nv_flat_27")]; + tensor var_5901 = mul(x = var_5695, y = var_1194)[name = string("op_5901")]; + tensor var_5902 = mul(x = nk_flat_27, y = update_mask_1)[name = string("op_5902")]; + tensor key_cache_57 = add(x = var_5901, y = var_5902)[name = string("key_cache_57")]; + tensor var_5908 = mul(x = var_5715, y = var_1194)[name = string("op_5908")]; + tensor var_5909 = mul(x = nv_flat_27, y = update_mask_1)[name = string("op_5909")]; + tensor value_cache_57 = add(x = var_5908, y = var_5909)[name = string("value_cache_57")]; + tensor kc_79_axes_0 = const()[name = string("kc_79_axes_0"), val = tensor([2])]; + tensor kc_79 = squeeze(axes = kc_79_axes_0, x = key_cache_57)[name = string("kc_79")]; + tensor var_5918 = const()[name = string("op_5918"), val = tensor([1, 8, 128, 256])]; + tensor kc_81 = reshape(shape = var_5918, x = kc_79)[name = string("kc_81")]; + tensor vc_79_axes_0 = const()[name = string("vc_79_axes_0"), val = tensor([2])]; + tensor vc_79 = squeeze(axes = vc_79_axes_0, x = value_cache_57)[name = string("vc_79")]; + tensor var_5926 = const()[name = string("op_5926"), val = tensor([1, 8, 128, 256])]; + tensor vc_81 = reshape(shape = var_5926, x = vc_79)[name = string("vc_81")]; + tensor var_5929_axes_0 = const()[name = string("op_5929_axes_0"), val = tensor([2])]; + tensor var_5929 = expand_dims(axes = var_5929_axes_0, x = kc_81)[name = string("op_5929")]; + tensor var_5937_reps_0 = const()[name = string("op_5937_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_5937 = tile(reps = var_5937_reps_0, x = var_5929)[name = string("op_5937")]; + tensor var_5942 = const()[name = string("op_5942"), val = tensor([1, 16, 128, 256])]; + tensor kc_83 = reshape(shape = var_5942, x = var_5937)[name = string("kc_83")]; + tensor var_5945_axes_0 = const()[name = string("op_5945_axes_0"), val = tensor([2])]; + tensor var_5945 = expand_dims(axes = var_5945_axes_0, x = vc_81)[name = string("op_5945")]; + tensor var_5953_reps_0 = const()[name = string("op_5953_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_5953 = tile(reps = var_5953_reps_0, x = var_5945)[name = string("op_5953")]; + tensor var_5958 = const()[name = string("op_5958"), val = tensor([1, 16, 128, 256])]; + tensor vc_83 = reshape(shape = var_5958, x = var_5953)[name = string("vc_83")]; + bool var_5960_transpose_x_0 = const()[name = string("op_5960_transpose_x_0"), val = bool(false)]; + bool var_5960_transpose_y_0 = const()[name = string("op_5960_transpose_y_0"), val = bool(false)]; + tensor var_5960 = matmul(transpose_x = var_5960_transpose_x_0, transpose_y = var_5960_transpose_y_0, x = q_111, y = kc_83)[name = string("op_5960")]; + fp32 _inversed_attn_weights_105_y_0 = const()[name = string("_inversed_attn_weights_105_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_105 = mul(x = var_5960, y = _inversed_attn_weights_105_y_0)[name = string("_inversed_attn_weights_105")]; + tensor attn_weights_107 = add(x = _inversed_attn_weights_105, y = mask_1)[name = string("attn_weights_107")]; + int32 var_5974 = const()[name = string("op_5974"), val = int32(-1)]; + tensor attn_weights_111 = softmax(axis = var_5974, x = attn_weights_107)[name = string("attn_weights_111")]; + bool attn_output_53_transpose_x_1 = const()[name = string("attn_output_53_transpose_x_1"), val = bool(false)]; + bool attn_output_53_transpose_y_1 = const()[name = string("attn_output_53_transpose_y_1"), val = bool(true)]; + tensor attn_output_53 = matmul(transpose_x = attn_output_53_transpose_x_1, transpose_y = attn_output_53_transpose_y_1, x = attn_weights_111, y = vc_83)[name = string("attn_output_53")]; + tensor var_5983_perm_0 = const()[name = string("op_5983_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5987 = const()[name = string("op_5987"), val = tensor([1, 1, -1])]; + tensor var_5983 = transpose(perm = var_5983_perm_0, x = attn_output_53)[name = string("transpose_56")]; + tensor input_133 = reshape(shape = var_5987, x = var_5983)[name = string("input_133")]; + tensor attn_output_55 = linear(bias = linear_1_bias_0, weight = layers_13_self_attn_o_proj_weight, x = input_133)[name = string("linear_94")]; + tensor var_5993_axes_0 = const()[name = string("op_5993_axes_0"), val = tensor([0])]; + tensor var_5993 = squeeze(axes = var_5993_axes_0, x = attn_output_55)[name = string("op_5993")]; + tensor var_5995_axes_0 = const()[name = string("op_5995_axes_0"), val = tensor([0])]; + tensor var_5995 = squeeze(axes = var_5995_axes_0, x = var_5993)[name = string("op_5995")]; + tensor var_5997_axes_0 = const()[name = string("op_5997_axes_0"), val = tensor([-1])]; + tensor var_5997 = expand_dims(axes = var_5997_axes_0, x = var_5995)[name = string("op_5997")]; + tensor attn_4d_27_axes_0 = const()[name = string("attn_4d_27_axes_0"), val = tensor([-1])]; + tensor attn_4d_27 = expand_dims(axes = attn_4d_27_axes_0, x = var_5997)[name = string("attn_4d_27")]; + tensor hidden_53 = add(x = hidden_51, y = attn_4d_27)[name = string("hidden_53")]; + tensor var_6003_axes_0 = const()[name = string("op_6003_axes_0"), val = tensor([-1])]; + tensor var_6003 = squeeze(axes = var_6003_axes_0, x = hidden_53)[name = string("op_6003")]; + tensor var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor([-1])]; + tensor var_6005 = squeeze(axes = var_6005_axes_0, x = var_6003)[name = string("op_6005")]; + tensor hidden_states_331_axes_0 = const()[name = string("hidden_states_331_axes_0"), val = tensor([0])]; + tensor hidden_states_331 = expand_dims(axes = hidden_states_331_axes_0, x = var_6005)[name = string("hidden_states_331")]; + fp32 var_6011_promoted = const()[name = string("op_6011_promoted"), val = fp32(0x1p+1)]; + tensor var_6017 = pow(x = hidden_states_331, y = var_6011_promoted)[name = string("op_6017")]; + tensor variance_111_axes_0 = const()[name = string("variance_111_axes_0"), val = tensor([-1])]; + bool variance_111_keep_dims_0 = const()[name = string("variance_111_keep_dims_0"), val = bool(true)]; + tensor variance_111 = reduce_mean(axes = variance_111_axes_0, keep_dims = variance_111_keep_dims_0, x = var_6017)[name = string("variance_111")]; + fp32 var_6020 = const()[name = string("op_6020"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6021 = add(x = variance_111, y = var_6020)[name = string("op_6021")]; + fp32 var_6022_epsilon_0 = const()[name = string("op_6022_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6022 = rsqrt(epsilon = var_6022_epsilon_0, x = var_6021)[name = string("op_6022")]; + tensor hidden_states_335 = mul(x = hidden_states_331, y = var_6022)[name = string("hidden_states_335")]; + tensor const_140 = const()[name = string("const_140"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774356800)))]; + tensor input_135 = mul(x = const_140, y = hidden_states_335)[name = string("input_135")]; + tensor input_137 = linear(bias = linear_4_bias_0, weight = layers_13_mlp_gate_proj_weight, x = input_135)[name = string("linear_95")]; + tensor var_6032 = silu(x = input_137)[name = string("op_6032")]; + tensor var_6034 = linear(bias = linear_4_bias_0, weight = layers_13_mlp_up_proj_weight, x = input_135)[name = string("linear_96")]; + tensor input_139 = mul(x = var_6032, y = var_6034)[name = string("input_139")]; + tensor mlp_out_27 = linear(bias = linear_1_bias_0, weight = layers_13_mlp_down_proj_weight, x = input_139)[name = string("linear_97")]; + tensor var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor([0])]; + tensor var_6039 = squeeze(axes = var_6039_axes_0, x = mlp_out_27)[name = string("op_6039")]; + tensor var_6041_axes_0 = const()[name = string("op_6041_axes_0"), val = tensor([0])]; + tensor var_6041 = squeeze(axes = var_6041_axes_0, x = var_6039)[name = string("op_6041")]; + tensor var_6043_axes_0 = const()[name = string("op_6043_axes_0"), val = tensor([-1])]; + tensor var_6043 = expand_dims(axes = var_6043_axes_0, x = var_6041)[name = string("op_6043")]; + tensor mlp_4d_27_axes_0 = const()[name = string("mlp_4d_27_axes_0"), val = tensor([-1])]; + tensor mlp_4d_27 = expand_dims(axes = mlp_4d_27_axes_0, x = var_6043)[name = string("mlp_4d_27")]; + tensor hidden_55 = add(x = hidden_53, y = mlp_4d_27)[name = string("hidden_55")]; + tensor var_6057_begin_0 = const()[name = string("op_6057_begin_0"), val = tensor([0, 14336, 0, 0])]; + tensor var_6057_end_0 = const()[name = string("op_6057_end_0"), val = tensor([1, 15360, 1, 256])]; + tensor var_6057_end_mask_0 = const()[name = string("op_6057_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6057 = slice_by_index(begin = var_6057_begin_0, end = var_6057_end_0, end_mask = var_6057_end_mask_0, x = cast_3)[name = string("op_6057")]; + tensor var_6077_begin_0 = const()[name = string("op_6077_begin_0"), val = tensor([0, 14336, 0, 0])]; + tensor var_6077_end_0 = const()[name = string("op_6077_end_0"), val = tensor([1, 15360, 1, 256])]; + tensor var_6077_end_mask_0 = const()[name = string("op_6077_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6077 = slice_by_index(begin = var_6077_begin_0, end = var_6077_end_0, end_mask = var_6077_end_mask_0, x = cast_4)[name = string("op_6077")]; + tensor var_6089_axes_0 = const()[name = string("op_6089_axes_0"), val = tensor([-1])]; + tensor var_6089 = squeeze(axes = var_6089_axes_0, x = hidden_55)[name = string("op_6089")]; + tensor var_6091_axes_0 = const()[name = string("op_6091_axes_0"), val = tensor([-1])]; + tensor var_6091 = squeeze(axes = var_6091_axes_0, x = var_6089)[name = string("op_6091")]; + tensor hidden_states_337_axes_0 = const()[name = string("hidden_states_337_axes_0"), val = tensor([0])]; + tensor hidden_states_337 = expand_dims(axes = hidden_states_337_axes_0, x = var_6091)[name = string("hidden_states_337")]; + fp32 var_6097_promoted = const()[name = string("op_6097_promoted"), val = fp32(0x1p+1)]; + tensor var_6103 = pow(x = hidden_states_337, y = var_6097_promoted)[name = string("op_6103")]; + tensor variance_113_axes_0 = const()[name = string("variance_113_axes_0"), val = tensor([-1])]; + bool variance_113_keep_dims_0 = const()[name = string("variance_113_keep_dims_0"), val = bool(true)]; + tensor variance_113 = reduce_mean(axes = variance_113_axes_0, keep_dims = variance_113_keep_dims_0, x = var_6103)[name = string("variance_113")]; + fp32 var_6106 = const()[name = string("op_6106"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6107 = add(x = variance_113, y = var_6106)[name = string("op_6107")]; + fp32 var_6108_epsilon_0 = const()[name = string("op_6108_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6108 = rsqrt(epsilon = var_6108_epsilon_0, x = var_6107)[name = string("op_6108")]; + tensor hidden_states_341 = mul(x = hidden_states_337, y = var_6108)[name = string("hidden_states_341")]; + tensor const_141 = const()[name = string("const_141"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774360960)))]; + tensor input_141 = mul(x = const_141, y = hidden_states_341)[name = string("input_141")]; + tensor q_113 = linear(bias = linear_0_bias_0, weight = layers_14_self_attn_q_proj_weight, x = input_141)[name = string("linear_98")]; + tensor k_113 = linear(bias = linear_1_bias_0, weight = layers_14_self_attn_k_proj_weight, x = input_141)[name = string("linear_99")]; + tensor v_85 = linear(bias = linear_1_bias_0, weight = layers_14_self_attn_v_proj_weight, x = input_141)[name = string("linear_100")]; + tensor var_6125 = const()[name = string("op_6125"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_343 = reshape(shape = var_6125, x = q_113)[name = string("hidden_states_343")]; + tensor var_6131 = const()[name = string("op_6131"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_349 = reshape(shape = var_6131, x = k_113)[name = string("hidden_states_349")]; + tensor var_6137 = const()[name = string("op_6137"), val = tensor([1, 1, 8, 128])]; + tensor v_87 = reshape(shape = var_6137, x = v_85)[name = string("v_87")]; + fp32 var_6142_promoted = const()[name = string("op_6142_promoted"), val = fp32(0x1p+1)]; + tensor var_6148 = pow(x = hidden_states_343, y = var_6142_promoted)[name = string("op_6148")]; + tensor variance_115_axes_0 = const()[name = string("variance_115_axes_0"), val = tensor([-1])]; + bool variance_115_keep_dims_0 = const()[name = string("variance_115_keep_dims_0"), val = bool(true)]; + tensor variance_115 = reduce_mean(axes = variance_115_axes_0, keep_dims = variance_115_keep_dims_0, x = var_6148)[name = string("variance_115")]; + fp32 var_6151 = const()[name = string("op_6151"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6152 = add(x = variance_115, y = var_6151)[name = string("op_6152")]; + fp32 var_6153_epsilon_0 = const()[name = string("op_6153_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6153 = rsqrt(epsilon = var_6153_epsilon_0, x = var_6152)[name = string("op_6153")]; + tensor hidden_states_347 = mul(x = hidden_states_343, y = var_6153)[name = string("hidden_states_347")]; + tensor const_142 = const()[name = string("const_142"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774365120)))]; + tensor q_115 = mul(x = const_142, y = hidden_states_347)[name = string("q_115")]; + fp32 var_6160_promoted = const()[name = string("op_6160_promoted"), val = fp32(0x1p+1)]; + tensor var_6166 = pow(x = hidden_states_349, y = var_6160_promoted)[name = string("op_6166")]; + tensor variance_117_axes_0 = const()[name = string("variance_117_axes_0"), val = tensor([-1])]; + bool variance_117_keep_dims_0 = const()[name = string("variance_117_keep_dims_0"), val = bool(true)]; + tensor variance_117 = reduce_mean(axes = variance_117_axes_0, keep_dims = variance_117_keep_dims_0, x = var_6166)[name = string("variance_117")]; + fp32 var_6169 = const()[name = string("op_6169"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6170 = add(x = variance_117, y = var_6169)[name = string("op_6170")]; + fp32 var_6171_epsilon_0 = const()[name = string("op_6171_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6171 = rsqrt(epsilon = var_6171_epsilon_0, x = var_6170)[name = string("op_6171")]; + tensor hidden_states_353 = mul(x = hidden_states_349, y = var_6171)[name = string("hidden_states_353")]; + tensor const_143 = const()[name = string("const_143"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774365696)))]; + tensor k_115 = mul(x = const_143, y = hidden_states_353)[name = string("k_115")]; + tensor q_117_perm_0 = const()[name = string("q_117_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_117_perm_0 = const()[name = string("k_117_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_89_perm_0 = const()[name = string("v_89_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_117 = transpose(perm = q_117_perm_0, x = q_115)[name = string("transpose_55")]; + tensor var_6188 = mul(x = q_117, y = cos_3)[name = string("op_6188")]; + tensor x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_57 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_117)[name = string("x1_57")]; + tensor x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_57 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_117)[name = string("x2_57")]; + fp32 const_146_promoted = const()[name = string("const_146_promoted"), val = fp32(-0x1p+0)]; + tensor var_6209 = mul(x = x2_57, y = const_146_promoted)[name = string("op_6209")]; + int32 var_6211 = const()[name = string("op_6211"), val = int32(-1)]; + bool var_6212_interleave_0 = const()[name = string("op_6212_interleave_0"), val = bool(false)]; + tensor var_6212 = concat(axis = var_6211, interleave = var_6212_interleave_0, values = (var_6209, x1_57))[name = string("op_6212")]; + tensor var_6213 = mul(x = var_6212, y = sin_3)[name = string("op_6213")]; + tensor q_119 = add(x = var_6188, y = var_6213)[name = string("q_119")]; + tensor k_117 = transpose(perm = k_117_perm_0, x = k_115)[name = string("transpose_54")]; + tensor var_6216 = mul(x = k_117, y = cos_3)[name = string("op_6216")]; + tensor x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_59 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_117)[name = string("x1_59")]; + tensor x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_59 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_117)[name = string("x2_59")]; + fp32 const_149_promoted = const()[name = string("const_149_promoted"), val = fp32(-0x1p+0)]; + tensor var_6237 = mul(x = x2_59, y = const_149_promoted)[name = string("op_6237")]; + int32 var_6239 = const()[name = string("op_6239"), val = int32(-1)]; + bool var_6240_interleave_0 = const()[name = string("op_6240_interleave_0"), val = bool(false)]; + tensor var_6240 = concat(axis = var_6239, interleave = var_6240_interleave_0, values = (var_6237, x1_59))[name = string("op_6240")]; + tensor var_6241 = mul(x = var_6240, y = sin_3)[name = string("op_6241")]; + tensor k_119 = add(x = var_6216, y = var_6241)[name = string("k_119")]; + tensor var_6248 = const()[name = string("op_6248"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_29 = reshape(shape = var_6248, x = k_119)[name = string("nk_flat_29")]; + tensor var_6254 = const()[name = string("op_6254"), val = tensor([1, 1024, 1, 1])]; + tensor v_89 = transpose(perm = v_89_perm_0, x = v_87)[name = string("transpose_53")]; + tensor nv_flat_29 = reshape(shape = var_6254, x = v_89)[name = string("nv_flat_29")]; + tensor var_6263 = mul(x = var_6057, y = var_1194)[name = string("op_6263")]; + tensor var_6264 = mul(x = nk_flat_29, y = update_mask_1)[name = string("op_6264")]; + tensor key_cache_61 = add(x = var_6263, y = var_6264)[name = string("key_cache_61")]; + tensor var_6270 = mul(x = var_6077, y = var_1194)[name = string("op_6270")]; + tensor var_6271 = mul(x = nv_flat_29, y = update_mask_1)[name = string("op_6271")]; + tensor value_cache_61 = add(x = var_6270, y = var_6271)[name = string("value_cache_61")]; + tensor kc_85_axes_0 = const()[name = string("kc_85_axes_0"), val = tensor([2])]; + tensor kc_85 = squeeze(axes = kc_85_axes_0, x = key_cache_61)[name = string("kc_85")]; + tensor var_6280 = const()[name = string("op_6280"), val = tensor([1, 8, 128, 256])]; + tensor kc_87 = reshape(shape = var_6280, x = kc_85)[name = string("kc_87")]; + tensor vc_85_axes_0 = const()[name = string("vc_85_axes_0"), val = tensor([2])]; + tensor vc_85 = squeeze(axes = vc_85_axes_0, x = value_cache_61)[name = string("vc_85")]; + tensor var_6288 = const()[name = string("op_6288"), val = tensor([1, 8, 128, 256])]; + tensor vc_87 = reshape(shape = var_6288, x = vc_85)[name = string("vc_87")]; + tensor var_6291_axes_0 = const()[name = string("op_6291_axes_0"), val = tensor([2])]; + tensor var_6291 = expand_dims(axes = var_6291_axes_0, x = kc_87)[name = string("op_6291")]; + tensor var_6299_reps_0 = const()[name = string("op_6299_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_6299 = tile(reps = var_6299_reps_0, x = var_6291)[name = string("op_6299")]; + tensor var_6304 = const()[name = string("op_6304"), val = tensor([1, 16, 128, 256])]; + tensor kc_89 = reshape(shape = var_6304, x = var_6299)[name = string("kc_89")]; + tensor var_6307_axes_0 = const()[name = string("op_6307_axes_0"), val = tensor([2])]; + tensor var_6307 = expand_dims(axes = var_6307_axes_0, x = vc_87)[name = string("op_6307")]; + tensor var_6315_reps_0 = const()[name = string("op_6315_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_6315 = tile(reps = var_6315_reps_0, x = var_6307)[name = string("op_6315")]; + tensor var_6320 = const()[name = string("op_6320"), val = tensor([1, 16, 128, 256])]; + tensor vc_89 = reshape(shape = var_6320, x = var_6315)[name = string("vc_89")]; + bool var_6322_transpose_x_0 = const()[name = string("op_6322_transpose_x_0"), val = bool(false)]; + bool var_6322_transpose_y_0 = const()[name = string("op_6322_transpose_y_0"), val = bool(false)]; + tensor var_6322 = matmul(transpose_x = var_6322_transpose_x_0, transpose_y = var_6322_transpose_y_0, x = q_119, y = kc_89)[name = string("op_6322")]; + fp32 _inversed_attn_weights_113_y_0 = const()[name = string("_inversed_attn_weights_113_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_113 = mul(x = var_6322, y = _inversed_attn_weights_113_y_0)[name = string("_inversed_attn_weights_113")]; + tensor attn_weights_115 = add(x = _inversed_attn_weights_113, y = mask_1)[name = string("attn_weights_115")]; + int32 var_6336 = const()[name = string("op_6336"), val = int32(-1)]; + tensor attn_weights_119 = softmax(axis = var_6336, x = attn_weights_115)[name = string("attn_weights_119")]; + bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(false)]; + bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(true)]; + tensor attn_output_57 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_119, y = vc_89)[name = string("attn_output_57")]; + tensor var_6345_perm_0 = const()[name = string("op_6345_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_6349 = const()[name = string("op_6349"), val = tensor([1, 1, -1])]; + tensor var_6345 = transpose(perm = var_6345_perm_0, x = attn_output_57)[name = string("transpose_52")]; + tensor input_143 = reshape(shape = var_6349, x = var_6345)[name = string("input_143")]; + tensor attn_output_59 = linear(bias = linear_1_bias_0, weight = layers_14_self_attn_o_proj_weight, x = input_143)[name = string("linear_101")]; + tensor var_6355_axes_0 = const()[name = string("op_6355_axes_0"), val = tensor([0])]; + tensor var_6355 = squeeze(axes = var_6355_axes_0, x = attn_output_59)[name = string("op_6355")]; + tensor var_6357_axes_0 = const()[name = string("op_6357_axes_0"), val = tensor([0])]; + tensor var_6357 = squeeze(axes = var_6357_axes_0, x = var_6355)[name = string("op_6357")]; + tensor var_6359_axes_0 = const()[name = string("op_6359_axes_0"), val = tensor([-1])]; + tensor var_6359 = expand_dims(axes = var_6359_axes_0, x = var_6357)[name = string("op_6359")]; + tensor attn_4d_29_axes_0 = const()[name = string("attn_4d_29_axes_0"), val = tensor([-1])]; + tensor attn_4d_29 = expand_dims(axes = attn_4d_29_axes_0, x = var_6359)[name = string("attn_4d_29")]; + tensor hidden_57 = add(x = hidden_55, y = attn_4d_29)[name = string("hidden_57")]; + tensor var_6365_axes_0 = const()[name = string("op_6365_axes_0"), val = tensor([-1])]; + tensor var_6365 = squeeze(axes = var_6365_axes_0, x = hidden_57)[name = string("op_6365")]; + tensor var_6367_axes_0 = const()[name = string("op_6367_axes_0"), val = tensor([-1])]; + tensor var_6367 = squeeze(axes = var_6367_axes_0, x = var_6365)[name = string("op_6367")]; + tensor hidden_states_355_axes_0 = const()[name = string("hidden_states_355_axes_0"), val = tensor([0])]; + tensor hidden_states_355 = expand_dims(axes = hidden_states_355_axes_0, x = var_6367)[name = string("hidden_states_355")]; + fp32 var_6373_promoted = const()[name = string("op_6373_promoted"), val = fp32(0x1p+1)]; + tensor var_6379 = pow(x = hidden_states_355, y = var_6373_promoted)[name = string("op_6379")]; + tensor variance_119_axes_0 = const()[name = string("variance_119_axes_0"), val = tensor([-1])]; + bool variance_119_keep_dims_0 = const()[name = string("variance_119_keep_dims_0"), val = bool(true)]; + tensor variance_119 = reduce_mean(axes = variance_119_axes_0, keep_dims = variance_119_keep_dims_0, x = var_6379)[name = string("variance_119")]; + fp32 var_6382 = const()[name = string("op_6382"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6383 = add(x = variance_119, y = var_6382)[name = string("op_6383")]; + fp32 var_6384_epsilon_0 = const()[name = string("op_6384_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6384 = rsqrt(epsilon = var_6384_epsilon_0, x = var_6383)[name = string("op_6384")]; + tensor hidden_states_359 = mul(x = hidden_states_355, y = var_6384)[name = string("hidden_states_359")]; + tensor const_150 = const()[name = string("const_150"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774366272)))]; + tensor input_145 = mul(x = const_150, y = hidden_states_359)[name = string("input_145")]; + tensor input_147 = linear(bias = linear_4_bias_0, weight = layers_14_mlp_gate_proj_weight, x = input_145)[name = string("linear_102")]; + tensor var_6394 = silu(x = input_147)[name = string("op_6394")]; + tensor var_6396 = linear(bias = linear_4_bias_0, weight = layers_14_mlp_up_proj_weight, x = input_145)[name = string("linear_103")]; + tensor input_149 = mul(x = var_6394, y = var_6396)[name = string("input_149")]; + tensor mlp_out_29 = linear(bias = linear_1_bias_0, weight = layers_14_mlp_down_proj_weight, x = input_149)[name = string("linear_104")]; + tensor var_6401_axes_0 = const()[name = string("op_6401_axes_0"), val = tensor([0])]; + tensor var_6401 = squeeze(axes = var_6401_axes_0, x = mlp_out_29)[name = string("op_6401")]; + tensor var_6403_axes_0 = const()[name = string("op_6403_axes_0"), val = tensor([0])]; + tensor var_6403 = squeeze(axes = var_6403_axes_0, x = var_6401)[name = string("op_6403")]; + tensor var_6405_axes_0 = const()[name = string("op_6405_axes_0"), val = tensor([-1])]; + tensor var_6405 = expand_dims(axes = var_6405_axes_0, x = var_6403)[name = string("op_6405")]; + tensor mlp_4d_29_axes_0 = const()[name = string("mlp_4d_29_axes_0"), val = tensor([-1])]; + tensor mlp_4d_29 = expand_dims(axes = mlp_4d_29_axes_0, x = var_6405)[name = string("mlp_4d_29")]; + tensor hidden_59 = add(x = hidden_57, y = mlp_4d_29)[name = string("hidden_59")]; + tensor var_6419_begin_0 = const()[name = string("op_6419_begin_0"), val = tensor([0, 15360, 0, 0])]; + tensor var_6419_end_0 = const()[name = string("op_6419_end_0"), val = tensor([1, 16384, 1, 256])]; + tensor var_6419_end_mask_0 = const()[name = string("op_6419_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6419 = slice_by_index(begin = var_6419_begin_0, end = var_6419_end_0, end_mask = var_6419_end_mask_0, x = cast_3)[name = string("op_6419")]; + tensor var_6439_begin_0 = const()[name = string("op_6439_begin_0"), val = tensor([0, 15360, 0, 0])]; + tensor var_6439_end_0 = const()[name = string("op_6439_end_0"), val = tensor([1, 16384, 1, 256])]; + tensor var_6439_end_mask_0 = const()[name = string("op_6439_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6439 = slice_by_index(begin = var_6439_begin_0, end = var_6439_end_0, end_mask = var_6439_end_mask_0, x = cast_4)[name = string("op_6439")]; + tensor var_6451_axes_0 = const()[name = string("op_6451_axes_0"), val = tensor([-1])]; + tensor var_6451 = squeeze(axes = var_6451_axes_0, x = hidden_59)[name = string("op_6451")]; + tensor var_6453_axes_0 = const()[name = string("op_6453_axes_0"), val = tensor([-1])]; + tensor var_6453 = squeeze(axes = var_6453_axes_0, x = var_6451)[name = string("op_6453")]; + tensor hidden_states_361_axes_0 = const()[name = string("hidden_states_361_axes_0"), val = tensor([0])]; + tensor hidden_states_361 = expand_dims(axes = hidden_states_361_axes_0, x = var_6453)[name = string("hidden_states_361")]; + fp32 var_6459_promoted = const()[name = string("op_6459_promoted"), val = fp32(0x1p+1)]; + tensor var_6465 = pow(x = hidden_states_361, y = var_6459_promoted)[name = string("op_6465")]; + tensor variance_121_axes_0 = const()[name = string("variance_121_axes_0"), val = tensor([-1])]; + bool variance_121_keep_dims_0 = const()[name = string("variance_121_keep_dims_0"), val = bool(true)]; + tensor variance_121 = reduce_mean(axes = variance_121_axes_0, keep_dims = variance_121_keep_dims_0, x = var_6465)[name = string("variance_121")]; + fp32 var_6468 = const()[name = string("op_6468"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6469 = add(x = variance_121, y = var_6468)[name = string("op_6469")]; + fp32 var_6470_epsilon_0 = const()[name = string("op_6470_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6470 = rsqrt(epsilon = var_6470_epsilon_0, x = var_6469)[name = string("op_6470")]; + tensor hidden_states_365 = mul(x = hidden_states_361, y = var_6470)[name = string("hidden_states_365")]; + tensor const_151 = const()[name = string("const_151"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774370432)))]; + tensor input_151 = mul(x = const_151, y = hidden_states_365)[name = string("input_151")]; + tensor q_121 = linear(bias = linear_0_bias_0, weight = layers_15_self_attn_q_proj_weight, x = input_151)[name = string("linear_105")]; + tensor k_121 = linear(bias = linear_1_bias_0, weight = layers_15_self_attn_k_proj_weight, x = input_151)[name = string("linear_106")]; + tensor v_91 = linear(bias = linear_1_bias_0, weight = layers_15_self_attn_v_proj_weight, x = input_151)[name = string("linear_107")]; + tensor var_6487 = const()[name = string("op_6487"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_367 = reshape(shape = var_6487, x = q_121)[name = string("hidden_states_367")]; + tensor var_6493 = const()[name = string("op_6493"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_373 = reshape(shape = var_6493, x = k_121)[name = string("hidden_states_373")]; + tensor var_6499 = const()[name = string("op_6499"), val = tensor([1, 1, 8, 128])]; + tensor v_93 = reshape(shape = var_6499, x = v_91)[name = string("v_93")]; + fp32 var_6504_promoted = const()[name = string("op_6504_promoted"), val = fp32(0x1p+1)]; + tensor var_6510 = pow(x = hidden_states_367, y = var_6504_promoted)[name = string("op_6510")]; + tensor variance_123_axes_0 = const()[name = string("variance_123_axes_0"), val = tensor([-1])]; + bool variance_123_keep_dims_0 = const()[name = string("variance_123_keep_dims_0"), val = bool(true)]; + tensor variance_123 = reduce_mean(axes = variance_123_axes_0, keep_dims = variance_123_keep_dims_0, x = var_6510)[name = string("variance_123")]; + fp32 var_6513 = const()[name = string("op_6513"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6514 = add(x = variance_123, y = var_6513)[name = string("op_6514")]; + fp32 var_6515_epsilon_0 = const()[name = string("op_6515_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6515 = rsqrt(epsilon = var_6515_epsilon_0, x = var_6514)[name = string("op_6515")]; + tensor hidden_states_371 = mul(x = hidden_states_367, y = var_6515)[name = string("hidden_states_371")]; + tensor const_152 = const()[name = string("const_152"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774374592)))]; + tensor q_123 = mul(x = const_152, y = hidden_states_371)[name = string("q_123")]; + fp32 var_6522_promoted = const()[name = string("op_6522_promoted"), val = fp32(0x1p+1)]; + tensor var_6528 = pow(x = hidden_states_373, y = var_6522_promoted)[name = string("op_6528")]; + tensor variance_125_axes_0 = const()[name = string("variance_125_axes_0"), val = tensor([-1])]; + bool variance_125_keep_dims_0 = const()[name = string("variance_125_keep_dims_0"), val = bool(true)]; + tensor variance_125 = reduce_mean(axes = variance_125_axes_0, keep_dims = variance_125_keep_dims_0, x = var_6528)[name = string("variance_125")]; + fp32 var_6531 = const()[name = string("op_6531"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6532 = add(x = variance_125, y = var_6531)[name = string("op_6532")]; + fp32 var_6533_epsilon_0 = const()[name = string("op_6533_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6533 = rsqrt(epsilon = var_6533_epsilon_0, x = var_6532)[name = string("op_6533")]; + tensor hidden_states_377 = mul(x = hidden_states_373, y = var_6533)[name = string("hidden_states_377")]; + tensor const_153 = const()[name = string("const_153"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774375168)))]; + tensor k_123 = mul(x = const_153, y = hidden_states_377)[name = string("k_123")]; + tensor q_125_perm_0 = const()[name = string("q_125_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_125_perm_0 = const()[name = string("k_125_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_95_perm_0 = const()[name = string("v_95_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_125 = transpose(perm = q_125_perm_0, x = q_123)[name = string("transpose_51")]; + tensor var_6550 = mul(x = q_125, y = cos_3)[name = string("op_6550")]; + tensor x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_61 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_125)[name = string("x1_61")]; + tensor x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_61 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_125)[name = string("x2_61")]; + fp32 const_156_promoted = const()[name = string("const_156_promoted"), val = fp32(-0x1p+0)]; + tensor var_6571 = mul(x = x2_61, y = const_156_promoted)[name = string("op_6571")]; + int32 var_6573 = const()[name = string("op_6573"), val = int32(-1)]; + bool var_6574_interleave_0 = const()[name = string("op_6574_interleave_0"), val = bool(false)]; + tensor var_6574 = concat(axis = var_6573, interleave = var_6574_interleave_0, values = (var_6571, x1_61))[name = string("op_6574")]; + tensor var_6575 = mul(x = var_6574, y = sin_3)[name = string("op_6575")]; + tensor q_127 = add(x = var_6550, y = var_6575)[name = string("q_127")]; + tensor k_125 = transpose(perm = k_125_perm_0, x = k_123)[name = string("transpose_50")]; + tensor var_6578 = mul(x = k_125, y = cos_3)[name = string("op_6578")]; + tensor x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_63 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_125)[name = string("x1_63")]; + tensor x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_63 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_125)[name = string("x2_63")]; + fp32 const_159_promoted = const()[name = string("const_159_promoted"), val = fp32(-0x1p+0)]; + tensor var_6599 = mul(x = x2_63, y = const_159_promoted)[name = string("op_6599")]; + int32 var_6601 = const()[name = string("op_6601"), val = int32(-1)]; + bool var_6602_interleave_0 = const()[name = string("op_6602_interleave_0"), val = bool(false)]; + tensor var_6602 = concat(axis = var_6601, interleave = var_6602_interleave_0, values = (var_6599, x1_63))[name = string("op_6602")]; + tensor var_6603 = mul(x = var_6602, y = sin_3)[name = string("op_6603")]; + tensor k_127 = add(x = var_6578, y = var_6603)[name = string("k_127")]; + tensor var_6610 = const()[name = string("op_6610"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_31 = reshape(shape = var_6610, x = k_127)[name = string("nk_flat_31")]; + tensor var_6616 = const()[name = string("op_6616"), val = tensor([1, 1024, 1, 1])]; + tensor v_95 = transpose(perm = v_95_perm_0, x = v_93)[name = string("transpose_49")]; + tensor nv_flat_31 = reshape(shape = var_6616, x = v_95)[name = string("nv_flat_31")]; + tensor var_6625 = mul(x = var_6419, y = var_1194)[name = string("op_6625")]; + tensor var_6626 = mul(x = nk_flat_31, y = update_mask_1)[name = string("op_6626")]; + tensor key_cache_65 = add(x = var_6625, y = var_6626)[name = string("key_cache_65")]; + tensor var_6632 = mul(x = var_6439, y = var_1194)[name = string("op_6632")]; + tensor var_6633 = mul(x = nv_flat_31, y = update_mask_1)[name = string("op_6633")]; + tensor value_cache_65 = add(x = var_6632, y = var_6633)[name = string("value_cache_65")]; + tensor kc_91_axes_0 = const()[name = string("kc_91_axes_0"), val = tensor([2])]; + tensor kc_91 = squeeze(axes = kc_91_axes_0, x = key_cache_65)[name = string("kc_91")]; + tensor var_6642 = const()[name = string("op_6642"), val = tensor([1, 8, 128, 256])]; + tensor kc_93 = reshape(shape = var_6642, x = kc_91)[name = string("kc_93")]; + tensor vc_91_axes_0 = const()[name = string("vc_91_axes_0"), val = tensor([2])]; + tensor vc_91 = squeeze(axes = vc_91_axes_0, x = value_cache_65)[name = string("vc_91")]; + tensor var_6650 = const()[name = string("op_6650"), val = tensor([1, 8, 128, 256])]; + tensor vc_93 = reshape(shape = var_6650, x = vc_91)[name = string("vc_93")]; + tensor var_6653_axes_0 = const()[name = string("op_6653_axes_0"), val = tensor([2])]; + tensor var_6653 = expand_dims(axes = var_6653_axes_0, x = kc_93)[name = string("op_6653")]; + tensor var_6661_reps_0 = const()[name = string("op_6661_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_6661 = tile(reps = var_6661_reps_0, x = var_6653)[name = string("op_6661")]; + tensor var_6666 = const()[name = string("op_6666"), val = tensor([1, 16, 128, 256])]; + tensor kc_95 = reshape(shape = var_6666, x = var_6661)[name = string("kc_95")]; + tensor var_6669_axes_0 = const()[name = string("op_6669_axes_0"), val = tensor([2])]; + tensor var_6669 = expand_dims(axes = var_6669_axes_0, x = vc_93)[name = string("op_6669")]; + tensor var_6677_reps_0 = const()[name = string("op_6677_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_6677 = tile(reps = var_6677_reps_0, x = var_6669)[name = string("op_6677")]; + tensor var_6682 = const()[name = string("op_6682"), val = tensor([1, 16, 128, 256])]; + tensor vc_95 = reshape(shape = var_6682, x = var_6677)[name = string("vc_95")]; + bool var_6684_transpose_x_0 = const()[name = string("op_6684_transpose_x_0"), val = bool(false)]; + bool var_6684_transpose_y_0 = const()[name = string("op_6684_transpose_y_0"), val = bool(false)]; + tensor var_6684 = matmul(transpose_x = var_6684_transpose_x_0, transpose_y = var_6684_transpose_y_0, x = q_127, y = kc_95)[name = string("op_6684")]; + fp32 _inversed_attn_weights_121_y_0 = const()[name = string("_inversed_attn_weights_121_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_121 = mul(x = var_6684, y = _inversed_attn_weights_121_y_0)[name = string("_inversed_attn_weights_121")]; + tensor attn_weights_123 = add(x = _inversed_attn_weights_121, y = mask_1)[name = string("attn_weights_123")]; + int32 var_6698 = const()[name = string("op_6698"), val = int32(-1)]; + tensor attn_weights_127 = softmax(axis = var_6698, x = attn_weights_123)[name = string("attn_weights_127")]; + bool attn_output_61_transpose_x_1 = const()[name = string("attn_output_61_transpose_x_1"), val = bool(false)]; + bool attn_output_61_transpose_y_1 = const()[name = string("attn_output_61_transpose_y_1"), val = bool(true)]; + tensor attn_output_61 = matmul(transpose_x = attn_output_61_transpose_x_1, transpose_y = attn_output_61_transpose_y_1, x = attn_weights_127, y = vc_95)[name = string("attn_output_61")]; + tensor var_6707_perm_0 = const()[name = string("op_6707_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_6711 = const()[name = string("op_6711"), val = tensor([1, 1, -1])]; + tensor var_6707 = transpose(perm = var_6707_perm_0, x = attn_output_61)[name = string("transpose_48")]; + tensor input_153 = reshape(shape = var_6711, x = var_6707)[name = string("input_153")]; + tensor attn_output_63 = linear(bias = linear_1_bias_0, weight = layers_15_self_attn_o_proj_weight, x = input_153)[name = string("linear_108")]; + tensor var_6717_axes_0 = const()[name = string("op_6717_axes_0"), val = tensor([0])]; + tensor var_6717 = squeeze(axes = var_6717_axes_0, x = attn_output_63)[name = string("op_6717")]; + tensor var_6719_axes_0 = const()[name = string("op_6719_axes_0"), val = tensor([0])]; + tensor var_6719 = squeeze(axes = var_6719_axes_0, x = var_6717)[name = string("op_6719")]; + tensor var_6721_axes_0 = const()[name = string("op_6721_axes_0"), val = tensor([-1])]; + tensor var_6721 = expand_dims(axes = var_6721_axes_0, x = var_6719)[name = string("op_6721")]; + tensor attn_4d_31_axes_0 = const()[name = string("attn_4d_31_axes_0"), val = tensor([-1])]; + tensor attn_4d_31 = expand_dims(axes = attn_4d_31_axes_0, x = var_6721)[name = string("attn_4d_31")]; + tensor hidden_61 = add(x = hidden_59, y = attn_4d_31)[name = string("hidden_61")]; + tensor var_6727_axes_0 = const()[name = string("op_6727_axes_0"), val = tensor([-1])]; + tensor var_6727 = squeeze(axes = var_6727_axes_0, x = hidden_61)[name = string("op_6727")]; + tensor var_6729_axes_0 = const()[name = string("op_6729_axes_0"), val = tensor([-1])]; + tensor var_6729 = squeeze(axes = var_6729_axes_0, x = var_6727)[name = string("op_6729")]; + tensor hidden_states_379_axes_0 = const()[name = string("hidden_states_379_axes_0"), val = tensor([0])]; + tensor hidden_states_379 = expand_dims(axes = hidden_states_379_axes_0, x = var_6729)[name = string("hidden_states_379")]; + fp32 var_6735_promoted = const()[name = string("op_6735_promoted"), val = fp32(0x1p+1)]; + tensor var_6741 = pow(x = hidden_states_379, y = var_6735_promoted)[name = string("op_6741")]; + tensor variance_127_axes_0 = const()[name = string("variance_127_axes_0"), val = tensor([-1])]; + bool variance_127_keep_dims_0 = const()[name = string("variance_127_keep_dims_0"), val = bool(true)]; + tensor variance_127 = reduce_mean(axes = variance_127_axes_0, keep_dims = variance_127_keep_dims_0, x = var_6741)[name = string("variance_127")]; + fp32 var_6744 = const()[name = string("op_6744"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6745 = add(x = variance_127, y = var_6744)[name = string("op_6745")]; + fp32 var_6746_epsilon_0 = const()[name = string("op_6746_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6746 = rsqrt(epsilon = var_6746_epsilon_0, x = var_6745)[name = string("op_6746")]; + tensor hidden_states_383 = mul(x = hidden_states_379, y = var_6746)[name = string("hidden_states_383")]; + tensor const_160 = const()[name = string("const_160"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774375744)))]; + tensor input_155 = mul(x = const_160, y = hidden_states_383)[name = string("input_155")]; + tensor input_157 = linear(bias = linear_4_bias_0, weight = layers_15_mlp_gate_proj_weight, x = input_155)[name = string("linear_109")]; + tensor var_6756 = silu(x = input_157)[name = string("op_6756")]; + tensor var_6758 = linear(bias = linear_4_bias_0, weight = layers_15_mlp_up_proj_weight, x = input_155)[name = string("linear_110")]; + tensor input_159 = mul(x = var_6756, y = var_6758)[name = string("input_159")]; + tensor mlp_out_31 = linear(bias = linear_1_bias_0, weight = layers_15_mlp_down_proj_weight, x = input_159)[name = string("linear_111")]; + tensor var_6763_axes_0 = const()[name = string("op_6763_axes_0"), val = tensor([0])]; + tensor var_6763 = squeeze(axes = var_6763_axes_0, x = mlp_out_31)[name = string("op_6763")]; + tensor var_6765_axes_0 = const()[name = string("op_6765_axes_0"), val = tensor([0])]; + tensor var_6765 = squeeze(axes = var_6765_axes_0, x = var_6763)[name = string("op_6765")]; + tensor var_6767_axes_0 = const()[name = string("op_6767_axes_0"), val = tensor([-1])]; + tensor var_6767 = expand_dims(axes = var_6767_axes_0, x = var_6765)[name = string("op_6767")]; + tensor mlp_4d_31_axes_0 = const()[name = string("mlp_4d_31_axes_0"), val = tensor([-1])]; + tensor mlp_4d_31 = expand_dims(axes = mlp_4d_31_axes_0, x = var_6767)[name = string("mlp_4d_31")]; + tensor hidden_63 = add(x = hidden_61, y = mlp_4d_31)[name = string("hidden_63")]; + tensor var_6781_begin_0 = const()[name = string("op_6781_begin_0"), val = tensor([0, 16384, 0, 0])]; + tensor var_6781_end_0 = const()[name = string("op_6781_end_0"), val = tensor([1, 17408, 1, 256])]; + tensor var_6781_end_mask_0 = const()[name = string("op_6781_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6781 = slice_by_index(begin = var_6781_begin_0, end = var_6781_end_0, end_mask = var_6781_end_mask_0, x = cast_3)[name = string("op_6781")]; + tensor var_6801_begin_0 = const()[name = string("op_6801_begin_0"), val = tensor([0, 16384, 0, 0])]; + tensor var_6801_end_0 = const()[name = string("op_6801_end_0"), val = tensor([1, 17408, 1, 256])]; + tensor var_6801_end_mask_0 = const()[name = string("op_6801_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_6801 = slice_by_index(begin = var_6801_begin_0, end = var_6801_end_0, end_mask = var_6801_end_mask_0, x = cast_4)[name = string("op_6801")]; + tensor var_6813_axes_0 = const()[name = string("op_6813_axes_0"), val = tensor([-1])]; + tensor var_6813 = squeeze(axes = var_6813_axes_0, x = hidden_63)[name = string("op_6813")]; + tensor var_6815_axes_0 = const()[name = string("op_6815_axes_0"), val = tensor([-1])]; + tensor var_6815 = squeeze(axes = var_6815_axes_0, x = var_6813)[name = string("op_6815")]; + tensor hidden_states_385_axes_0 = const()[name = string("hidden_states_385_axes_0"), val = tensor([0])]; + tensor hidden_states_385 = expand_dims(axes = hidden_states_385_axes_0, x = var_6815)[name = string("hidden_states_385")]; + fp32 var_6821_promoted = const()[name = string("op_6821_promoted"), val = fp32(0x1p+1)]; + tensor var_6827 = pow(x = hidden_states_385, y = var_6821_promoted)[name = string("op_6827")]; + tensor variance_129_axes_0 = const()[name = string("variance_129_axes_0"), val = tensor([-1])]; + bool variance_129_keep_dims_0 = const()[name = string("variance_129_keep_dims_0"), val = bool(true)]; + tensor variance_129 = reduce_mean(axes = variance_129_axes_0, keep_dims = variance_129_keep_dims_0, x = var_6827)[name = string("variance_129")]; + fp32 var_6830 = const()[name = string("op_6830"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6831 = add(x = variance_129, y = var_6830)[name = string("op_6831")]; + fp32 var_6832_epsilon_0 = const()[name = string("op_6832_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6832 = rsqrt(epsilon = var_6832_epsilon_0, x = var_6831)[name = string("op_6832")]; + tensor hidden_states_389 = mul(x = hidden_states_385, y = var_6832)[name = string("hidden_states_389")]; + tensor const_161 = const()[name = string("const_161"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774379904)))]; + tensor input_161 = mul(x = const_161, y = hidden_states_389)[name = string("input_161")]; + tensor q_129 = linear(bias = linear_0_bias_0, weight = layers_16_self_attn_q_proj_weight, x = input_161)[name = string("linear_112")]; + tensor k_129 = linear(bias = linear_1_bias_0, weight = layers_16_self_attn_k_proj_weight, x = input_161)[name = string("linear_113")]; + tensor v_97 = linear(bias = linear_1_bias_0, weight = layers_16_self_attn_v_proj_weight, x = input_161)[name = string("linear_114")]; + tensor var_6849 = const()[name = string("op_6849"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_391 = reshape(shape = var_6849, x = q_129)[name = string("hidden_states_391")]; + tensor var_6855 = const()[name = string("op_6855"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_397 = reshape(shape = var_6855, x = k_129)[name = string("hidden_states_397")]; + tensor var_6861 = const()[name = string("op_6861"), val = tensor([1, 1, 8, 128])]; + tensor v_99 = reshape(shape = var_6861, x = v_97)[name = string("v_99")]; + fp32 var_6866_promoted = const()[name = string("op_6866_promoted"), val = fp32(0x1p+1)]; + tensor var_6872 = pow(x = hidden_states_391, y = var_6866_promoted)[name = string("op_6872")]; + tensor variance_131_axes_0 = const()[name = string("variance_131_axes_0"), val = tensor([-1])]; + bool variance_131_keep_dims_0 = const()[name = string("variance_131_keep_dims_0"), val = bool(true)]; + tensor variance_131 = reduce_mean(axes = variance_131_axes_0, keep_dims = variance_131_keep_dims_0, x = var_6872)[name = string("variance_131")]; + fp32 var_6875 = const()[name = string("op_6875"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6876 = add(x = variance_131, y = var_6875)[name = string("op_6876")]; + fp32 var_6877_epsilon_0 = const()[name = string("op_6877_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6877 = rsqrt(epsilon = var_6877_epsilon_0, x = var_6876)[name = string("op_6877")]; + tensor hidden_states_395 = mul(x = hidden_states_391, y = var_6877)[name = string("hidden_states_395")]; + tensor const_162 = const()[name = string("const_162"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774384064)))]; + tensor q_131 = mul(x = const_162, y = hidden_states_395)[name = string("q_131")]; + fp32 var_6884_promoted = const()[name = string("op_6884_promoted"), val = fp32(0x1p+1)]; + tensor var_6890 = pow(x = hidden_states_397, y = var_6884_promoted)[name = string("op_6890")]; + tensor variance_133_axes_0 = const()[name = string("variance_133_axes_0"), val = tensor([-1])]; + bool variance_133_keep_dims_0 = const()[name = string("variance_133_keep_dims_0"), val = bool(true)]; + tensor variance_133 = reduce_mean(axes = variance_133_axes_0, keep_dims = variance_133_keep_dims_0, x = var_6890)[name = string("variance_133")]; + fp32 var_6893 = const()[name = string("op_6893"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_6894 = add(x = variance_133, y = var_6893)[name = string("op_6894")]; + fp32 var_6895_epsilon_0 = const()[name = string("op_6895_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_6895 = rsqrt(epsilon = var_6895_epsilon_0, x = var_6894)[name = string("op_6895")]; + tensor hidden_states_401 = mul(x = hidden_states_397, y = var_6895)[name = string("hidden_states_401")]; + tensor const_163 = const()[name = string("const_163"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774384640)))]; + tensor k_131 = mul(x = const_163, y = hidden_states_401)[name = string("k_131")]; + tensor q_133_perm_0 = const()[name = string("q_133_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_133_perm_0 = const()[name = string("k_133_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_101_perm_0 = const()[name = string("v_101_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_133 = transpose(perm = q_133_perm_0, x = q_131)[name = string("transpose_47")]; + tensor var_6912 = mul(x = q_133, y = cos_3)[name = string("op_6912")]; + tensor x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_65 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_133)[name = string("x1_65")]; + tensor x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_65 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_133)[name = string("x2_65")]; + fp32 const_166_promoted = const()[name = string("const_166_promoted"), val = fp32(-0x1p+0)]; + tensor var_6933 = mul(x = x2_65, y = const_166_promoted)[name = string("op_6933")]; + int32 var_6935 = const()[name = string("op_6935"), val = int32(-1)]; + bool var_6936_interleave_0 = const()[name = string("op_6936_interleave_0"), val = bool(false)]; + tensor var_6936 = concat(axis = var_6935, interleave = var_6936_interleave_0, values = (var_6933, x1_65))[name = string("op_6936")]; + tensor var_6937 = mul(x = var_6936, y = sin_3)[name = string("op_6937")]; + tensor q_135 = add(x = var_6912, y = var_6937)[name = string("q_135")]; + tensor k_133 = transpose(perm = k_133_perm_0, x = k_131)[name = string("transpose_46")]; + tensor var_6940 = mul(x = k_133, y = cos_3)[name = string("op_6940")]; + tensor x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_67 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_133)[name = string("x1_67")]; + tensor x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_67 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_133)[name = string("x2_67")]; + fp32 const_169_promoted = const()[name = string("const_169_promoted"), val = fp32(-0x1p+0)]; + tensor var_6961 = mul(x = x2_67, y = const_169_promoted)[name = string("op_6961")]; + int32 var_6963 = const()[name = string("op_6963"), val = int32(-1)]; + bool var_6964_interleave_0 = const()[name = string("op_6964_interleave_0"), val = bool(false)]; + tensor var_6964 = concat(axis = var_6963, interleave = var_6964_interleave_0, values = (var_6961, x1_67))[name = string("op_6964")]; + tensor var_6965 = mul(x = var_6964, y = sin_3)[name = string("op_6965")]; + tensor k_135 = add(x = var_6940, y = var_6965)[name = string("k_135")]; + tensor var_6972 = const()[name = string("op_6972"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_33 = reshape(shape = var_6972, x = k_135)[name = string("nk_flat_33")]; + tensor var_6978 = const()[name = string("op_6978"), val = tensor([1, 1024, 1, 1])]; + tensor v_101 = transpose(perm = v_101_perm_0, x = v_99)[name = string("transpose_45")]; + tensor nv_flat_33 = reshape(shape = var_6978, x = v_101)[name = string("nv_flat_33")]; + tensor var_6987 = mul(x = var_6781, y = var_1194)[name = string("op_6987")]; + tensor var_6988 = mul(x = nk_flat_33, y = update_mask_1)[name = string("op_6988")]; + tensor key_cache_69 = add(x = var_6987, y = var_6988)[name = string("key_cache_69")]; + tensor var_6994 = mul(x = var_6801, y = var_1194)[name = string("op_6994")]; + tensor var_6995 = mul(x = nv_flat_33, y = update_mask_1)[name = string("op_6995")]; + tensor value_cache_69 = add(x = var_6994, y = var_6995)[name = string("value_cache_69")]; + tensor kc_97_axes_0 = const()[name = string("kc_97_axes_0"), val = tensor([2])]; + tensor kc_97 = squeeze(axes = kc_97_axes_0, x = key_cache_69)[name = string("kc_97")]; + tensor var_7004 = const()[name = string("op_7004"), val = tensor([1, 8, 128, 256])]; + tensor kc_99 = reshape(shape = var_7004, x = kc_97)[name = string("kc_99")]; + tensor vc_97_axes_0 = const()[name = string("vc_97_axes_0"), val = tensor([2])]; + tensor vc_97 = squeeze(axes = vc_97_axes_0, x = value_cache_69)[name = string("vc_97")]; + tensor var_7012 = const()[name = string("op_7012"), val = tensor([1, 8, 128, 256])]; + tensor vc_99 = reshape(shape = var_7012, x = vc_97)[name = string("vc_99")]; + tensor var_7015_axes_0 = const()[name = string("op_7015_axes_0"), val = tensor([2])]; + tensor var_7015 = expand_dims(axes = var_7015_axes_0, x = kc_99)[name = string("op_7015")]; + tensor var_7023_reps_0 = const()[name = string("op_7023_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_7023 = tile(reps = var_7023_reps_0, x = var_7015)[name = string("op_7023")]; + tensor var_7028 = const()[name = string("op_7028"), val = tensor([1, 16, 128, 256])]; + tensor kc_101 = reshape(shape = var_7028, x = var_7023)[name = string("kc_101")]; + tensor var_7031_axes_0 = const()[name = string("op_7031_axes_0"), val = tensor([2])]; + tensor var_7031 = expand_dims(axes = var_7031_axes_0, x = vc_99)[name = string("op_7031")]; + tensor var_7039_reps_0 = const()[name = string("op_7039_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_7039 = tile(reps = var_7039_reps_0, x = var_7031)[name = string("op_7039")]; + tensor var_7044 = const()[name = string("op_7044"), val = tensor([1, 16, 128, 256])]; + tensor vc_101 = reshape(shape = var_7044, x = var_7039)[name = string("vc_101")]; + bool var_7046_transpose_x_0 = const()[name = string("op_7046_transpose_x_0"), val = bool(false)]; + bool var_7046_transpose_y_0 = const()[name = string("op_7046_transpose_y_0"), val = bool(false)]; + tensor var_7046 = matmul(transpose_x = var_7046_transpose_x_0, transpose_y = var_7046_transpose_y_0, x = q_135, y = kc_101)[name = string("op_7046")]; + fp32 _inversed_attn_weights_129_y_0 = const()[name = string("_inversed_attn_weights_129_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_129 = mul(x = var_7046, y = _inversed_attn_weights_129_y_0)[name = string("_inversed_attn_weights_129")]; + tensor attn_weights_131 = add(x = _inversed_attn_weights_129, y = mask_1)[name = string("attn_weights_131")]; + int32 var_7060 = const()[name = string("op_7060"), val = int32(-1)]; + tensor attn_weights_135 = softmax(axis = var_7060, x = attn_weights_131)[name = string("attn_weights_135")]; + bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(false)]; + bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(true)]; + tensor attn_output_65 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_135, y = vc_101)[name = string("attn_output_65")]; + tensor var_7069_perm_0 = const()[name = string("op_7069_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_7073 = const()[name = string("op_7073"), val = tensor([1, 1, -1])]; + tensor var_7069 = transpose(perm = var_7069_perm_0, x = attn_output_65)[name = string("transpose_44")]; + tensor input_163 = reshape(shape = var_7073, x = var_7069)[name = string("input_163")]; + tensor attn_output_67 = linear(bias = linear_1_bias_0, weight = layers_16_self_attn_o_proj_weight, x = input_163)[name = string("linear_115")]; + tensor var_7079_axes_0 = const()[name = string("op_7079_axes_0"), val = tensor([0])]; + tensor var_7079 = squeeze(axes = var_7079_axes_0, x = attn_output_67)[name = string("op_7079")]; + tensor var_7081_axes_0 = const()[name = string("op_7081_axes_0"), val = tensor([0])]; + tensor var_7081 = squeeze(axes = var_7081_axes_0, x = var_7079)[name = string("op_7081")]; + tensor var_7083_axes_0 = const()[name = string("op_7083_axes_0"), val = tensor([-1])]; + tensor var_7083 = expand_dims(axes = var_7083_axes_0, x = var_7081)[name = string("op_7083")]; + tensor attn_4d_33_axes_0 = const()[name = string("attn_4d_33_axes_0"), val = tensor([-1])]; + tensor attn_4d_33 = expand_dims(axes = attn_4d_33_axes_0, x = var_7083)[name = string("attn_4d_33")]; + tensor hidden_65 = add(x = hidden_63, y = attn_4d_33)[name = string("hidden_65")]; + tensor var_7089_axes_0 = const()[name = string("op_7089_axes_0"), val = tensor([-1])]; + tensor var_7089 = squeeze(axes = var_7089_axes_0, x = hidden_65)[name = string("op_7089")]; + tensor var_7091_axes_0 = const()[name = string("op_7091_axes_0"), val = tensor([-1])]; + tensor var_7091 = squeeze(axes = var_7091_axes_0, x = var_7089)[name = string("op_7091")]; + tensor hidden_states_403_axes_0 = const()[name = string("hidden_states_403_axes_0"), val = tensor([0])]; + tensor hidden_states_403 = expand_dims(axes = hidden_states_403_axes_0, x = var_7091)[name = string("hidden_states_403")]; + fp32 var_7097_promoted = const()[name = string("op_7097_promoted"), val = fp32(0x1p+1)]; + tensor var_7103 = pow(x = hidden_states_403, y = var_7097_promoted)[name = string("op_7103")]; + tensor variance_135_axes_0 = const()[name = string("variance_135_axes_0"), val = tensor([-1])]; + bool variance_135_keep_dims_0 = const()[name = string("variance_135_keep_dims_0"), val = bool(true)]; + tensor variance_135 = reduce_mean(axes = variance_135_axes_0, keep_dims = variance_135_keep_dims_0, x = var_7103)[name = string("variance_135")]; + fp32 var_7106 = const()[name = string("op_7106"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7107 = add(x = variance_135, y = var_7106)[name = string("op_7107")]; + fp32 var_7108_epsilon_0 = const()[name = string("op_7108_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7108 = rsqrt(epsilon = var_7108_epsilon_0, x = var_7107)[name = string("op_7108")]; + tensor hidden_states_407 = mul(x = hidden_states_403, y = var_7108)[name = string("hidden_states_407")]; + tensor const_170 = const()[name = string("const_170"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774385216)))]; + tensor input_165 = mul(x = const_170, y = hidden_states_407)[name = string("input_165")]; + tensor input_167 = linear(bias = linear_4_bias_0, weight = layers_16_mlp_gate_proj_weight, x = input_165)[name = string("linear_116")]; + tensor var_7118 = silu(x = input_167)[name = string("op_7118")]; + tensor var_7120 = linear(bias = linear_4_bias_0, weight = layers_16_mlp_up_proj_weight, x = input_165)[name = string("linear_117")]; + tensor input_169 = mul(x = var_7118, y = var_7120)[name = string("input_169")]; + tensor mlp_out_33 = linear(bias = linear_1_bias_0, weight = layers_16_mlp_down_proj_weight, x = input_169)[name = string("linear_118")]; + tensor var_7125_axes_0 = const()[name = string("op_7125_axes_0"), val = tensor([0])]; + tensor var_7125 = squeeze(axes = var_7125_axes_0, x = mlp_out_33)[name = string("op_7125")]; + tensor var_7127_axes_0 = const()[name = string("op_7127_axes_0"), val = tensor([0])]; + tensor var_7127 = squeeze(axes = var_7127_axes_0, x = var_7125)[name = string("op_7127")]; + tensor var_7129_axes_0 = const()[name = string("op_7129_axes_0"), val = tensor([-1])]; + tensor var_7129 = expand_dims(axes = var_7129_axes_0, x = var_7127)[name = string("op_7129")]; + tensor mlp_4d_33_axes_0 = const()[name = string("mlp_4d_33_axes_0"), val = tensor([-1])]; + tensor mlp_4d_33 = expand_dims(axes = mlp_4d_33_axes_0, x = var_7129)[name = string("mlp_4d_33")]; + tensor hidden_67 = add(x = hidden_65, y = mlp_4d_33)[name = string("hidden_67")]; + tensor var_7143_begin_0 = const()[name = string("op_7143_begin_0"), val = tensor([0, 17408, 0, 0])]; + tensor var_7143_end_0 = const()[name = string("op_7143_end_0"), val = tensor([1, 18432, 1, 256])]; + tensor var_7143_end_mask_0 = const()[name = string("op_7143_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7143 = slice_by_index(begin = var_7143_begin_0, end = var_7143_end_0, end_mask = var_7143_end_mask_0, x = cast_3)[name = string("op_7143")]; + tensor var_7163_begin_0 = const()[name = string("op_7163_begin_0"), val = tensor([0, 17408, 0, 0])]; + tensor var_7163_end_0 = const()[name = string("op_7163_end_0"), val = tensor([1, 18432, 1, 256])]; + tensor var_7163_end_mask_0 = const()[name = string("op_7163_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7163 = slice_by_index(begin = var_7163_begin_0, end = var_7163_end_0, end_mask = var_7163_end_mask_0, x = cast_4)[name = string("op_7163")]; + tensor var_7175_axes_0 = const()[name = string("op_7175_axes_0"), val = tensor([-1])]; + tensor var_7175 = squeeze(axes = var_7175_axes_0, x = hidden_67)[name = string("op_7175")]; + tensor var_7177_axes_0 = const()[name = string("op_7177_axes_0"), val = tensor([-1])]; + tensor var_7177 = squeeze(axes = var_7177_axes_0, x = var_7175)[name = string("op_7177")]; + tensor hidden_states_409_axes_0 = const()[name = string("hidden_states_409_axes_0"), val = tensor([0])]; + tensor hidden_states_409 = expand_dims(axes = hidden_states_409_axes_0, x = var_7177)[name = string("hidden_states_409")]; + fp32 var_7183_promoted = const()[name = string("op_7183_promoted"), val = fp32(0x1p+1)]; + tensor var_7189 = pow(x = hidden_states_409, y = var_7183_promoted)[name = string("op_7189")]; + tensor variance_137_axes_0 = const()[name = string("variance_137_axes_0"), val = tensor([-1])]; + bool variance_137_keep_dims_0 = const()[name = string("variance_137_keep_dims_0"), val = bool(true)]; + tensor variance_137 = reduce_mean(axes = variance_137_axes_0, keep_dims = variance_137_keep_dims_0, x = var_7189)[name = string("variance_137")]; + fp32 var_7192 = const()[name = string("op_7192"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7193 = add(x = variance_137, y = var_7192)[name = string("op_7193")]; + fp32 var_7194_epsilon_0 = const()[name = string("op_7194_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7194 = rsqrt(epsilon = var_7194_epsilon_0, x = var_7193)[name = string("op_7194")]; + tensor hidden_states_413 = mul(x = hidden_states_409, y = var_7194)[name = string("hidden_states_413")]; + tensor const_171 = const()[name = string("const_171"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774389376)))]; + tensor input_171 = mul(x = const_171, y = hidden_states_413)[name = string("input_171")]; + tensor q_137 = linear(bias = linear_0_bias_0, weight = layers_17_self_attn_q_proj_weight, x = input_171)[name = string("linear_119")]; + tensor k_137 = linear(bias = linear_1_bias_0, weight = layers_17_self_attn_k_proj_weight, x = input_171)[name = string("linear_120")]; + tensor v_103 = linear(bias = linear_1_bias_0, weight = layers_17_self_attn_v_proj_weight, x = input_171)[name = string("linear_121")]; + tensor var_7211 = const()[name = string("op_7211"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_415 = reshape(shape = var_7211, x = q_137)[name = string("hidden_states_415")]; + tensor var_7217 = const()[name = string("op_7217"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_421 = reshape(shape = var_7217, x = k_137)[name = string("hidden_states_421")]; + tensor var_7223 = const()[name = string("op_7223"), val = tensor([1, 1, 8, 128])]; + tensor v_105 = reshape(shape = var_7223, x = v_103)[name = string("v_105")]; + fp32 var_7228_promoted = const()[name = string("op_7228_promoted"), val = fp32(0x1p+1)]; + tensor var_7234 = pow(x = hidden_states_415, y = var_7228_promoted)[name = string("op_7234")]; + tensor variance_139_axes_0 = const()[name = string("variance_139_axes_0"), val = tensor([-1])]; + bool variance_139_keep_dims_0 = const()[name = string("variance_139_keep_dims_0"), val = bool(true)]; + tensor variance_139 = reduce_mean(axes = variance_139_axes_0, keep_dims = variance_139_keep_dims_0, x = var_7234)[name = string("variance_139")]; + fp32 var_7237 = const()[name = string("op_7237"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7238 = add(x = variance_139, y = var_7237)[name = string("op_7238")]; + fp32 var_7239_epsilon_0 = const()[name = string("op_7239_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7239 = rsqrt(epsilon = var_7239_epsilon_0, x = var_7238)[name = string("op_7239")]; + tensor hidden_states_419 = mul(x = hidden_states_415, y = var_7239)[name = string("hidden_states_419")]; + tensor const_172 = const()[name = string("const_172"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774393536)))]; + tensor q_139 = mul(x = const_172, y = hidden_states_419)[name = string("q_139")]; + fp32 var_7246_promoted = const()[name = string("op_7246_promoted"), val = fp32(0x1p+1)]; + tensor var_7252 = pow(x = hidden_states_421, y = var_7246_promoted)[name = string("op_7252")]; + tensor variance_141_axes_0 = const()[name = string("variance_141_axes_0"), val = tensor([-1])]; + bool variance_141_keep_dims_0 = const()[name = string("variance_141_keep_dims_0"), val = bool(true)]; + tensor variance_141 = reduce_mean(axes = variance_141_axes_0, keep_dims = variance_141_keep_dims_0, x = var_7252)[name = string("variance_141")]; + fp32 var_7255 = const()[name = string("op_7255"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7256 = add(x = variance_141, y = var_7255)[name = string("op_7256")]; + fp32 var_7257_epsilon_0 = const()[name = string("op_7257_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7257 = rsqrt(epsilon = var_7257_epsilon_0, x = var_7256)[name = string("op_7257")]; + tensor hidden_states_425 = mul(x = hidden_states_421, y = var_7257)[name = string("hidden_states_425")]; + tensor const_173 = const()[name = string("const_173"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774394112)))]; + tensor k_139 = mul(x = const_173, y = hidden_states_425)[name = string("k_139")]; + tensor q_141_perm_0 = const()[name = string("q_141_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_141_perm_0 = const()[name = string("k_141_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_107_perm_0 = const()[name = string("v_107_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_141 = transpose(perm = q_141_perm_0, x = q_139)[name = string("transpose_43")]; + tensor var_7274 = mul(x = q_141, y = cos_3)[name = string("op_7274")]; + tensor x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_69 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_141)[name = string("x1_69")]; + tensor x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_69 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_141)[name = string("x2_69")]; + fp32 const_176_promoted = const()[name = string("const_176_promoted"), val = fp32(-0x1p+0)]; + tensor var_7295 = mul(x = x2_69, y = const_176_promoted)[name = string("op_7295")]; + int32 var_7297 = const()[name = string("op_7297"), val = int32(-1)]; + bool var_7298_interleave_0 = const()[name = string("op_7298_interleave_0"), val = bool(false)]; + tensor var_7298 = concat(axis = var_7297, interleave = var_7298_interleave_0, values = (var_7295, x1_69))[name = string("op_7298")]; + tensor var_7299 = mul(x = var_7298, y = sin_3)[name = string("op_7299")]; + tensor q_143 = add(x = var_7274, y = var_7299)[name = string("q_143")]; + tensor k_141 = transpose(perm = k_141_perm_0, x = k_139)[name = string("transpose_42")]; + tensor var_7302 = mul(x = k_141, y = cos_3)[name = string("op_7302")]; + tensor x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_71 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_141)[name = string("x1_71")]; + tensor x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_71 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_141)[name = string("x2_71")]; + fp32 const_179_promoted = const()[name = string("const_179_promoted"), val = fp32(-0x1p+0)]; + tensor var_7323 = mul(x = x2_71, y = const_179_promoted)[name = string("op_7323")]; + int32 var_7325 = const()[name = string("op_7325"), val = int32(-1)]; + bool var_7326_interleave_0 = const()[name = string("op_7326_interleave_0"), val = bool(false)]; + tensor var_7326 = concat(axis = var_7325, interleave = var_7326_interleave_0, values = (var_7323, x1_71))[name = string("op_7326")]; + tensor var_7327 = mul(x = var_7326, y = sin_3)[name = string("op_7327")]; + tensor k_143 = add(x = var_7302, y = var_7327)[name = string("k_143")]; + tensor var_7334 = const()[name = string("op_7334"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_35 = reshape(shape = var_7334, x = k_143)[name = string("nk_flat_35")]; + tensor var_7340 = const()[name = string("op_7340"), val = tensor([1, 1024, 1, 1])]; + tensor v_107 = transpose(perm = v_107_perm_0, x = v_105)[name = string("transpose_41")]; + tensor nv_flat_35 = reshape(shape = var_7340, x = v_107)[name = string("nv_flat_35")]; + tensor var_7349 = mul(x = var_7143, y = var_1194)[name = string("op_7349")]; + tensor var_7350 = mul(x = nk_flat_35, y = update_mask_1)[name = string("op_7350")]; + tensor key_cache_73 = add(x = var_7349, y = var_7350)[name = string("key_cache_73")]; + tensor var_7356 = mul(x = var_7163, y = var_1194)[name = string("op_7356")]; + tensor var_7357 = mul(x = nv_flat_35, y = update_mask_1)[name = string("op_7357")]; + tensor value_cache_73 = add(x = var_7356, y = var_7357)[name = string("value_cache_73")]; + tensor kc_103_axes_0 = const()[name = string("kc_103_axes_0"), val = tensor([2])]; + tensor kc_103 = squeeze(axes = kc_103_axes_0, x = key_cache_73)[name = string("kc_103")]; + tensor var_7366 = const()[name = string("op_7366"), val = tensor([1, 8, 128, 256])]; + tensor kc_105 = reshape(shape = var_7366, x = kc_103)[name = string("kc_105")]; + tensor vc_103_axes_0 = const()[name = string("vc_103_axes_0"), val = tensor([2])]; + tensor vc_103 = squeeze(axes = vc_103_axes_0, x = value_cache_73)[name = string("vc_103")]; + tensor var_7374 = const()[name = string("op_7374"), val = tensor([1, 8, 128, 256])]; + tensor vc_105 = reshape(shape = var_7374, x = vc_103)[name = string("vc_105")]; + tensor var_7377_axes_0 = const()[name = string("op_7377_axes_0"), val = tensor([2])]; + tensor var_7377 = expand_dims(axes = var_7377_axes_0, x = kc_105)[name = string("op_7377")]; + tensor var_7385_reps_0 = const()[name = string("op_7385_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_7385 = tile(reps = var_7385_reps_0, x = var_7377)[name = string("op_7385")]; + tensor var_7390 = const()[name = string("op_7390"), val = tensor([1, 16, 128, 256])]; + tensor kc_107 = reshape(shape = var_7390, x = var_7385)[name = string("kc_107")]; + tensor var_7393_axes_0 = const()[name = string("op_7393_axes_0"), val = tensor([2])]; + tensor var_7393 = expand_dims(axes = var_7393_axes_0, x = vc_105)[name = string("op_7393")]; + tensor var_7401_reps_0 = const()[name = string("op_7401_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_7401 = tile(reps = var_7401_reps_0, x = var_7393)[name = string("op_7401")]; + tensor var_7406 = const()[name = string("op_7406"), val = tensor([1, 16, 128, 256])]; + tensor vc_107 = reshape(shape = var_7406, x = var_7401)[name = string("vc_107")]; + bool var_7408_transpose_x_0 = const()[name = string("op_7408_transpose_x_0"), val = bool(false)]; + bool var_7408_transpose_y_0 = const()[name = string("op_7408_transpose_y_0"), val = bool(false)]; + tensor var_7408 = matmul(transpose_x = var_7408_transpose_x_0, transpose_y = var_7408_transpose_y_0, x = q_143, y = kc_107)[name = string("op_7408")]; + fp32 _inversed_attn_weights_137_y_0 = const()[name = string("_inversed_attn_weights_137_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_137 = mul(x = var_7408, y = _inversed_attn_weights_137_y_0)[name = string("_inversed_attn_weights_137")]; + tensor attn_weights_139 = add(x = _inversed_attn_weights_137, y = mask_1)[name = string("attn_weights_139")]; + int32 var_7422 = const()[name = string("op_7422"), val = int32(-1)]; + tensor attn_weights_143 = softmax(axis = var_7422, x = attn_weights_139)[name = string("attn_weights_143")]; + bool attn_output_69_transpose_x_1 = const()[name = string("attn_output_69_transpose_x_1"), val = bool(false)]; + bool attn_output_69_transpose_y_1 = const()[name = string("attn_output_69_transpose_y_1"), val = bool(true)]; + tensor attn_output_69 = matmul(transpose_x = attn_output_69_transpose_x_1, transpose_y = attn_output_69_transpose_y_1, x = attn_weights_143, y = vc_107)[name = string("attn_output_69")]; + tensor var_7431_perm_0 = const()[name = string("op_7431_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_7435 = const()[name = string("op_7435"), val = tensor([1, 1, -1])]; + tensor var_7431 = transpose(perm = var_7431_perm_0, x = attn_output_69)[name = string("transpose_40")]; + tensor input_173 = reshape(shape = var_7435, x = var_7431)[name = string("input_173")]; + tensor attn_output_71 = linear(bias = linear_1_bias_0, weight = layers_17_self_attn_o_proj_weight, x = input_173)[name = string("linear_122")]; + tensor var_7441_axes_0 = const()[name = string("op_7441_axes_0"), val = tensor([0])]; + tensor var_7441 = squeeze(axes = var_7441_axes_0, x = attn_output_71)[name = string("op_7441")]; + tensor var_7443_axes_0 = const()[name = string("op_7443_axes_0"), val = tensor([0])]; + tensor var_7443 = squeeze(axes = var_7443_axes_0, x = var_7441)[name = string("op_7443")]; + tensor var_7445_axes_0 = const()[name = string("op_7445_axes_0"), val = tensor([-1])]; + tensor var_7445 = expand_dims(axes = var_7445_axes_0, x = var_7443)[name = string("op_7445")]; + tensor attn_4d_35_axes_0 = const()[name = string("attn_4d_35_axes_0"), val = tensor([-1])]; + tensor attn_4d_35 = expand_dims(axes = attn_4d_35_axes_0, x = var_7445)[name = string("attn_4d_35")]; + tensor hidden_69 = add(x = hidden_67, y = attn_4d_35)[name = string("hidden_69")]; + tensor var_7451_axes_0 = const()[name = string("op_7451_axes_0"), val = tensor([-1])]; + tensor var_7451 = squeeze(axes = var_7451_axes_0, x = hidden_69)[name = string("op_7451")]; + tensor var_7453_axes_0 = const()[name = string("op_7453_axes_0"), val = tensor([-1])]; + tensor var_7453 = squeeze(axes = var_7453_axes_0, x = var_7451)[name = string("op_7453")]; + tensor hidden_states_427_axes_0 = const()[name = string("hidden_states_427_axes_0"), val = tensor([0])]; + tensor hidden_states_427 = expand_dims(axes = hidden_states_427_axes_0, x = var_7453)[name = string("hidden_states_427")]; + fp32 var_7459_promoted = const()[name = string("op_7459_promoted"), val = fp32(0x1p+1)]; + tensor var_7465 = pow(x = hidden_states_427, y = var_7459_promoted)[name = string("op_7465")]; + tensor variance_143_axes_0 = const()[name = string("variance_143_axes_0"), val = tensor([-1])]; + bool variance_143_keep_dims_0 = const()[name = string("variance_143_keep_dims_0"), val = bool(true)]; + tensor variance_143 = reduce_mean(axes = variance_143_axes_0, keep_dims = variance_143_keep_dims_0, x = var_7465)[name = string("variance_143")]; + fp32 var_7468 = const()[name = string("op_7468"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7469 = add(x = variance_143, y = var_7468)[name = string("op_7469")]; + fp32 var_7470_epsilon_0 = const()[name = string("op_7470_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7470 = rsqrt(epsilon = var_7470_epsilon_0, x = var_7469)[name = string("op_7470")]; + tensor hidden_states_431 = mul(x = hidden_states_427, y = var_7470)[name = string("hidden_states_431")]; + tensor const_180 = const()[name = string("const_180"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774394688)))]; + tensor input_175 = mul(x = const_180, y = hidden_states_431)[name = string("input_175")]; + tensor input_177 = linear(bias = linear_4_bias_0, weight = layers_17_mlp_gate_proj_weight, x = input_175)[name = string("linear_123")]; + tensor var_7480 = silu(x = input_177)[name = string("op_7480")]; + tensor var_7482 = linear(bias = linear_4_bias_0, weight = layers_17_mlp_up_proj_weight, x = input_175)[name = string("linear_124")]; + tensor input_179 = mul(x = var_7480, y = var_7482)[name = string("input_179")]; + tensor mlp_out_35 = linear(bias = linear_1_bias_0, weight = layers_17_mlp_down_proj_weight, x = input_179)[name = string("linear_125")]; + tensor var_7487_axes_0 = const()[name = string("op_7487_axes_0"), val = tensor([0])]; + tensor var_7487 = squeeze(axes = var_7487_axes_0, x = mlp_out_35)[name = string("op_7487")]; + tensor var_7489_axes_0 = const()[name = string("op_7489_axes_0"), val = tensor([0])]; + tensor var_7489 = squeeze(axes = var_7489_axes_0, x = var_7487)[name = string("op_7489")]; + tensor var_7491_axes_0 = const()[name = string("op_7491_axes_0"), val = tensor([-1])]; + tensor var_7491 = expand_dims(axes = var_7491_axes_0, x = var_7489)[name = string("op_7491")]; + tensor mlp_4d_35_axes_0 = const()[name = string("mlp_4d_35_axes_0"), val = tensor([-1])]; + tensor mlp_4d_35 = expand_dims(axes = mlp_4d_35_axes_0, x = var_7491)[name = string("mlp_4d_35")]; + tensor hidden_71 = add(x = hidden_69, y = mlp_4d_35)[name = string("hidden_71")]; + tensor var_7505_begin_0 = const()[name = string("op_7505_begin_0"), val = tensor([0, 18432, 0, 0])]; + tensor var_7505_end_0 = const()[name = string("op_7505_end_0"), val = tensor([1, 19456, 1, 256])]; + tensor var_7505_end_mask_0 = const()[name = string("op_7505_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7505 = slice_by_index(begin = var_7505_begin_0, end = var_7505_end_0, end_mask = var_7505_end_mask_0, x = cast_3)[name = string("op_7505")]; + tensor var_7525_begin_0 = const()[name = string("op_7525_begin_0"), val = tensor([0, 18432, 0, 0])]; + tensor var_7525_end_0 = const()[name = string("op_7525_end_0"), val = tensor([1, 19456, 1, 256])]; + tensor var_7525_end_mask_0 = const()[name = string("op_7525_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7525 = slice_by_index(begin = var_7525_begin_0, end = var_7525_end_0, end_mask = var_7525_end_mask_0, x = cast_4)[name = string("op_7525")]; + tensor var_7537_axes_0 = const()[name = string("op_7537_axes_0"), val = tensor([-1])]; + tensor var_7537 = squeeze(axes = var_7537_axes_0, x = hidden_71)[name = string("op_7537")]; + tensor var_7539_axes_0 = const()[name = string("op_7539_axes_0"), val = tensor([-1])]; + tensor var_7539 = squeeze(axes = var_7539_axes_0, x = var_7537)[name = string("op_7539")]; + tensor hidden_states_433_axes_0 = const()[name = string("hidden_states_433_axes_0"), val = tensor([0])]; + tensor hidden_states_433 = expand_dims(axes = hidden_states_433_axes_0, x = var_7539)[name = string("hidden_states_433")]; + fp32 var_7545_promoted = const()[name = string("op_7545_promoted"), val = fp32(0x1p+1)]; + tensor var_7551 = pow(x = hidden_states_433, y = var_7545_promoted)[name = string("op_7551")]; + tensor variance_145_axes_0 = const()[name = string("variance_145_axes_0"), val = tensor([-1])]; + bool variance_145_keep_dims_0 = const()[name = string("variance_145_keep_dims_0"), val = bool(true)]; + tensor variance_145 = reduce_mean(axes = variance_145_axes_0, keep_dims = variance_145_keep_dims_0, x = var_7551)[name = string("variance_145")]; + fp32 var_7554 = const()[name = string("op_7554"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7555 = add(x = variance_145, y = var_7554)[name = string("op_7555")]; + fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7556 = rsqrt(epsilon = var_7556_epsilon_0, x = var_7555)[name = string("op_7556")]; + tensor hidden_states_437 = mul(x = hidden_states_433, y = var_7556)[name = string("hidden_states_437")]; + tensor const_181 = const()[name = string("const_181"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774398848)))]; + tensor input_181 = mul(x = const_181, y = hidden_states_437)[name = string("input_181")]; + tensor q_145 = linear(bias = linear_0_bias_0, weight = layers_18_self_attn_q_proj_weight, x = input_181)[name = string("linear_126")]; + tensor k_145 = linear(bias = linear_1_bias_0, weight = layers_18_self_attn_k_proj_weight, x = input_181)[name = string("linear_127")]; + tensor v_109 = linear(bias = linear_1_bias_0, weight = layers_18_self_attn_v_proj_weight, x = input_181)[name = string("linear_128")]; + tensor var_7573 = const()[name = string("op_7573"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_439 = reshape(shape = var_7573, x = q_145)[name = string("hidden_states_439")]; + tensor var_7579 = const()[name = string("op_7579"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_445 = reshape(shape = var_7579, x = k_145)[name = string("hidden_states_445")]; + tensor var_7585 = const()[name = string("op_7585"), val = tensor([1, 1, 8, 128])]; + tensor v_111 = reshape(shape = var_7585, x = v_109)[name = string("v_111")]; + fp32 var_7590_promoted = const()[name = string("op_7590_promoted"), val = fp32(0x1p+1)]; + tensor var_7596 = pow(x = hidden_states_439, y = var_7590_promoted)[name = string("op_7596")]; + tensor variance_147_axes_0 = const()[name = string("variance_147_axes_0"), val = tensor([-1])]; + bool variance_147_keep_dims_0 = const()[name = string("variance_147_keep_dims_0"), val = bool(true)]; + tensor variance_147 = reduce_mean(axes = variance_147_axes_0, keep_dims = variance_147_keep_dims_0, x = var_7596)[name = string("variance_147")]; + fp32 var_7599 = const()[name = string("op_7599"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7600 = add(x = variance_147, y = var_7599)[name = string("op_7600")]; + fp32 var_7601_epsilon_0 = const()[name = string("op_7601_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7601 = rsqrt(epsilon = var_7601_epsilon_0, x = var_7600)[name = string("op_7601")]; + tensor hidden_states_443 = mul(x = hidden_states_439, y = var_7601)[name = string("hidden_states_443")]; + tensor const_182 = const()[name = string("const_182"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774403008)))]; + tensor q_147 = mul(x = const_182, y = hidden_states_443)[name = string("q_147")]; + fp32 var_7608_promoted = const()[name = string("op_7608_promoted"), val = fp32(0x1p+1)]; + tensor var_7614 = pow(x = hidden_states_445, y = var_7608_promoted)[name = string("op_7614")]; + tensor variance_149_axes_0 = const()[name = string("variance_149_axes_0"), val = tensor([-1])]; + bool variance_149_keep_dims_0 = const()[name = string("variance_149_keep_dims_0"), val = bool(true)]; + tensor variance_149 = reduce_mean(axes = variance_149_axes_0, keep_dims = variance_149_keep_dims_0, x = var_7614)[name = string("variance_149")]; + fp32 var_7617 = const()[name = string("op_7617"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7618 = add(x = variance_149, y = var_7617)[name = string("op_7618")]; + fp32 var_7619_epsilon_0 = const()[name = string("op_7619_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7619 = rsqrt(epsilon = var_7619_epsilon_0, x = var_7618)[name = string("op_7619")]; + tensor hidden_states_449 = mul(x = hidden_states_445, y = var_7619)[name = string("hidden_states_449")]; + tensor const_183 = const()[name = string("const_183"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774403584)))]; + tensor k_147 = mul(x = const_183, y = hidden_states_449)[name = string("k_147")]; + tensor q_149_perm_0 = const()[name = string("q_149_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_149_perm_0 = const()[name = string("k_149_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_113_perm_0 = const()[name = string("v_113_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_149 = transpose(perm = q_149_perm_0, x = q_147)[name = string("transpose_39")]; + tensor var_7636 = mul(x = q_149, y = cos_3)[name = string("op_7636")]; + tensor x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_73 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_149)[name = string("x1_73")]; + tensor x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_73 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_149)[name = string("x2_73")]; + fp32 const_186_promoted = const()[name = string("const_186_promoted"), val = fp32(-0x1p+0)]; + tensor var_7657 = mul(x = x2_73, y = const_186_promoted)[name = string("op_7657")]; + int32 var_7659 = const()[name = string("op_7659"), val = int32(-1)]; + bool var_7660_interleave_0 = const()[name = string("op_7660_interleave_0"), val = bool(false)]; + tensor var_7660 = concat(axis = var_7659, interleave = var_7660_interleave_0, values = (var_7657, x1_73))[name = string("op_7660")]; + tensor var_7661 = mul(x = var_7660, y = sin_3)[name = string("op_7661")]; + tensor q_151 = add(x = var_7636, y = var_7661)[name = string("q_151")]; + tensor k_149 = transpose(perm = k_149_perm_0, x = k_147)[name = string("transpose_38")]; + tensor var_7664 = mul(x = k_149, y = cos_3)[name = string("op_7664")]; + tensor x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_75 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_149)[name = string("x1_75")]; + tensor x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_75 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_149)[name = string("x2_75")]; + fp32 const_189_promoted = const()[name = string("const_189_promoted"), val = fp32(-0x1p+0)]; + tensor var_7685 = mul(x = x2_75, y = const_189_promoted)[name = string("op_7685")]; + int32 var_7687 = const()[name = string("op_7687"), val = int32(-1)]; + bool var_7688_interleave_0 = const()[name = string("op_7688_interleave_0"), val = bool(false)]; + tensor var_7688 = concat(axis = var_7687, interleave = var_7688_interleave_0, values = (var_7685, x1_75))[name = string("op_7688")]; + tensor var_7689 = mul(x = var_7688, y = sin_3)[name = string("op_7689")]; + tensor k_151 = add(x = var_7664, y = var_7689)[name = string("k_151")]; + tensor var_7696 = const()[name = string("op_7696"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_37 = reshape(shape = var_7696, x = k_151)[name = string("nk_flat_37")]; + tensor var_7702 = const()[name = string("op_7702"), val = tensor([1, 1024, 1, 1])]; + tensor v_113 = transpose(perm = v_113_perm_0, x = v_111)[name = string("transpose_37")]; + tensor nv_flat_37 = reshape(shape = var_7702, x = v_113)[name = string("nv_flat_37")]; + tensor var_7711 = mul(x = var_7505, y = var_1194)[name = string("op_7711")]; + tensor var_7712 = mul(x = nk_flat_37, y = update_mask_1)[name = string("op_7712")]; + tensor key_cache_77 = add(x = var_7711, y = var_7712)[name = string("key_cache_77")]; + tensor var_7718 = mul(x = var_7525, y = var_1194)[name = string("op_7718")]; + tensor var_7719 = mul(x = nv_flat_37, y = update_mask_1)[name = string("op_7719")]; + tensor value_cache_77 = add(x = var_7718, y = var_7719)[name = string("value_cache_77")]; + tensor kc_109_axes_0 = const()[name = string("kc_109_axes_0"), val = tensor([2])]; + tensor kc_109 = squeeze(axes = kc_109_axes_0, x = key_cache_77)[name = string("kc_109")]; + tensor var_7728 = const()[name = string("op_7728"), val = tensor([1, 8, 128, 256])]; + tensor kc_111 = reshape(shape = var_7728, x = kc_109)[name = string("kc_111")]; + tensor vc_109_axes_0 = const()[name = string("vc_109_axes_0"), val = tensor([2])]; + tensor vc_109 = squeeze(axes = vc_109_axes_0, x = value_cache_77)[name = string("vc_109")]; + tensor var_7736 = const()[name = string("op_7736"), val = tensor([1, 8, 128, 256])]; + tensor vc_111 = reshape(shape = var_7736, x = vc_109)[name = string("vc_111")]; + tensor var_7739_axes_0 = const()[name = string("op_7739_axes_0"), val = tensor([2])]; + tensor var_7739 = expand_dims(axes = var_7739_axes_0, x = kc_111)[name = string("op_7739")]; + tensor var_7747_reps_0 = const()[name = string("op_7747_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_7747 = tile(reps = var_7747_reps_0, x = var_7739)[name = string("op_7747")]; + tensor var_7752 = const()[name = string("op_7752"), val = tensor([1, 16, 128, 256])]; + tensor kc_113 = reshape(shape = var_7752, x = var_7747)[name = string("kc_113")]; + tensor var_7755_axes_0 = const()[name = string("op_7755_axes_0"), val = tensor([2])]; + tensor var_7755 = expand_dims(axes = var_7755_axes_0, x = vc_111)[name = string("op_7755")]; + tensor var_7763_reps_0 = const()[name = string("op_7763_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_7763 = tile(reps = var_7763_reps_0, x = var_7755)[name = string("op_7763")]; + tensor var_7768 = const()[name = string("op_7768"), val = tensor([1, 16, 128, 256])]; + tensor vc_113 = reshape(shape = var_7768, x = var_7763)[name = string("vc_113")]; + bool var_7770_transpose_x_0 = const()[name = string("op_7770_transpose_x_0"), val = bool(false)]; + bool var_7770_transpose_y_0 = const()[name = string("op_7770_transpose_y_0"), val = bool(false)]; + tensor var_7770 = matmul(transpose_x = var_7770_transpose_x_0, transpose_y = var_7770_transpose_y_0, x = q_151, y = kc_113)[name = string("op_7770")]; + fp32 _inversed_attn_weights_145_y_0 = const()[name = string("_inversed_attn_weights_145_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_145 = mul(x = var_7770, y = _inversed_attn_weights_145_y_0)[name = string("_inversed_attn_weights_145")]; + tensor attn_weights_147 = add(x = _inversed_attn_weights_145, y = mask_1)[name = string("attn_weights_147")]; + int32 var_7784 = const()[name = string("op_7784"), val = int32(-1)]; + tensor attn_weights_151 = softmax(axis = var_7784, x = attn_weights_147)[name = string("attn_weights_151")]; + bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(false)]; + bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(true)]; + tensor attn_output_73 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_151, y = vc_113)[name = string("attn_output_73")]; + tensor var_7793_perm_0 = const()[name = string("op_7793_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_7797 = const()[name = string("op_7797"), val = tensor([1, 1, -1])]; + tensor var_7793 = transpose(perm = var_7793_perm_0, x = attn_output_73)[name = string("transpose_36")]; + tensor input_183 = reshape(shape = var_7797, x = var_7793)[name = string("input_183")]; + tensor attn_output_75 = linear(bias = linear_1_bias_0, weight = layers_18_self_attn_o_proj_weight, x = input_183)[name = string("linear_129")]; + tensor var_7803_axes_0 = const()[name = string("op_7803_axes_0"), val = tensor([0])]; + tensor var_7803 = squeeze(axes = var_7803_axes_0, x = attn_output_75)[name = string("op_7803")]; + tensor var_7805_axes_0 = const()[name = string("op_7805_axes_0"), val = tensor([0])]; + tensor var_7805 = squeeze(axes = var_7805_axes_0, x = var_7803)[name = string("op_7805")]; + tensor var_7807_axes_0 = const()[name = string("op_7807_axes_0"), val = tensor([-1])]; + tensor var_7807 = expand_dims(axes = var_7807_axes_0, x = var_7805)[name = string("op_7807")]; + tensor attn_4d_37_axes_0 = const()[name = string("attn_4d_37_axes_0"), val = tensor([-1])]; + tensor attn_4d_37 = expand_dims(axes = attn_4d_37_axes_0, x = var_7807)[name = string("attn_4d_37")]; + tensor hidden_73 = add(x = hidden_71, y = attn_4d_37)[name = string("hidden_73")]; + tensor var_7813_axes_0 = const()[name = string("op_7813_axes_0"), val = tensor([-1])]; + tensor var_7813 = squeeze(axes = var_7813_axes_0, x = hidden_73)[name = string("op_7813")]; + tensor var_7815_axes_0 = const()[name = string("op_7815_axes_0"), val = tensor([-1])]; + tensor var_7815 = squeeze(axes = var_7815_axes_0, x = var_7813)[name = string("op_7815")]; + tensor hidden_states_451_axes_0 = const()[name = string("hidden_states_451_axes_0"), val = tensor([0])]; + tensor hidden_states_451 = expand_dims(axes = hidden_states_451_axes_0, x = var_7815)[name = string("hidden_states_451")]; + fp32 var_7821_promoted = const()[name = string("op_7821_promoted"), val = fp32(0x1p+1)]; + tensor var_7827 = pow(x = hidden_states_451, y = var_7821_promoted)[name = string("op_7827")]; + tensor variance_151_axes_0 = const()[name = string("variance_151_axes_0"), val = tensor([-1])]; + bool variance_151_keep_dims_0 = const()[name = string("variance_151_keep_dims_0"), val = bool(true)]; + tensor variance_151 = reduce_mean(axes = variance_151_axes_0, keep_dims = variance_151_keep_dims_0, x = var_7827)[name = string("variance_151")]; + fp32 var_7830 = const()[name = string("op_7830"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7831 = add(x = variance_151, y = var_7830)[name = string("op_7831")]; + fp32 var_7832_epsilon_0 = const()[name = string("op_7832_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7832 = rsqrt(epsilon = var_7832_epsilon_0, x = var_7831)[name = string("op_7832")]; + tensor hidden_states_455 = mul(x = hidden_states_451, y = var_7832)[name = string("hidden_states_455")]; + tensor const_190 = const()[name = string("const_190"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774404160)))]; + tensor input_185 = mul(x = const_190, y = hidden_states_455)[name = string("input_185")]; + tensor input_187 = linear(bias = linear_4_bias_0, weight = layers_18_mlp_gate_proj_weight, x = input_185)[name = string("linear_130")]; + tensor var_7842 = silu(x = input_187)[name = string("op_7842")]; + tensor var_7844 = linear(bias = linear_4_bias_0, weight = layers_18_mlp_up_proj_weight, x = input_185)[name = string("linear_131")]; + tensor input_189 = mul(x = var_7842, y = var_7844)[name = string("input_189")]; + tensor mlp_out_37 = linear(bias = linear_1_bias_0, weight = layers_18_mlp_down_proj_weight, x = input_189)[name = string("linear_132")]; + tensor var_7849_axes_0 = const()[name = string("op_7849_axes_0"), val = tensor([0])]; + tensor var_7849 = squeeze(axes = var_7849_axes_0, x = mlp_out_37)[name = string("op_7849")]; + tensor var_7851_axes_0 = const()[name = string("op_7851_axes_0"), val = tensor([0])]; + tensor var_7851 = squeeze(axes = var_7851_axes_0, x = var_7849)[name = string("op_7851")]; + tensor var_7853_axes_0 = const()[name = string("op_7853_axes_0"), val = tensor([-1])]; + tensor var_7853 = expand_dims(axes = var_7853_axes_0, x = var_7851)[name = string("op_7853")]; + tensor mlp_4d_37_axes_0 = const()[name = string("mlp_4d_37_axes_0"), val = tensor([-1])]; + tensor mlp_4d_37 = expand_dims(axes = mlp_4d_37_axes_0, x = var_7853)[name = string("mlp_4d_37")]; + tensor hidden_75 = add(x = hidden_73, y = mlp_4d_37)[name = string("hidden_75")]; + tensor var_7867_begin_0 = const()[name = string("op_7867_begin_0"), val = tensor([0, 19456, 0, 0])]; + tensor var_7867_end_0 = const()[name = string("op_7867_end_0"), val = tensor([1, 20480, 1, 256])]; + tensor var_7867_end_mask_0 = const()[name = string("op_7867_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7867 = slice_by_index(begin = var_7867_begin_0, end = var_7867_end_0, end_mask = var_7867_end_mask_0, x = cast_3)[name = string("op_7867")]; + tensor var_7887_begin_0 = const()[name = string("op_7887_begin_0"), val = tensor([0, 19456, 0, 0])]; + tensor var_7887_end_0 = const()[name = string("op_7887_end_0"), val = tensor([1, 20480, 1, 256])]; + tensor var_7887_end_mask_0 = const()[name = string("op_7887_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7887 = slice_by_index(begin = var_7887_begin_0, end = var_7887_end_0, end_mask = var_7887_end_mask_0, x = cast_4)[name = string("op_7887")]; + tensor var_7899_axes_0 = const()[name = string("op_7899_axes_0"), val = tensor([-1])]; + tensor var_7899 = squeeze(axes = var_7899_axes_0, x = hidden_75)[name = string("op_7899")]; + tensor var_7901_axes_0 = const()[name = string("op_7901_axes_0"), val = tensor([-1])]; + tensor var_7901 = squeeze(axes = var_7901_axes_0, x = var_7899)[name = string("op_7901")]; + tensor hidden_states_457_axes_0 = const()[name = string("hidden_states_457_axes_0"), val = tensor([0])]; + tensor hidden_states_457 = expand_dims(axes = hidden_states_457_axes_0, x = var_7901)[name = string("hidden_states_457")]; + fp32 var_7907_promoted = const()[name = string("op_7907_promoted"), val = fp32(0x1p+1)]; + tensor var_7913 = pow(x = hidden_states_457, y = var_7907_promoted)[name = string("op_7913")]; + tensor variance_153_axes_0 = const()[name = string("variance_153_axes_0"), val = tensor([-1])]; + bool variance_153_keep_dims_0 = const()[name = string("variance_153_keep_dims_0"), val = bool(true)]; + tensor variance_153 = reduce_mean(axes = variance_153_axes_0, keep_dims = variance_153_keep_dims_0, x = var_7913)[name = string("variance_153")]; + fp32 var_7916 = const()[name = string("op_7916"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7917 = add(x = variance_153, y = var_7916)[name = string("op_7917")]; + fp32 var_7918_epsilon_0 = const()[name = string("op_7918_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7918 = rsqrt(epsilon = var_7918_epsilon_0, x = var_7917)[name = string("op_7918")]; + tensor hidden_states_461 = mul(x = hidden_states_457, y = var_7918)[name = string("hidden_states_461")]; + tensor const_191 = const()[name = string("const_191"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774408320)))]; + tensor input_191 = mul(x = const_191, y = hidden_states_461)[name = string("input_191")]; + tensor q_153 = linear(bias = linear_0_bias_0, weight = layers_19_self_attn_q_proj_weight, x = input_191)[name = string("linear_133")]; + tensor k_153 = linear(bias = linear_1_bias_0, weight = layers_19_self_attn_k_proj_weight, x = input_191)[name = string("linear_134")]; + tensor v_115 = linear(bias = linear_1_bias_0, weight = layers_19_self_attn_v_proj_weight, x = input_191)[name = string("linear_135")]; + tensor var_7935 = const()[name = string("op_7935"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_463 = reshape(shape = var_7935, x = q_153)[name = string("hidden_states_463")]; + tensor var_7941 = const()[name = string("op_7941"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_469 = reshape(shape = var_7941, x = k_153)[name = string("hidden_states_469")]; + tensor var_7947 = const()[name = string("op_7947"), val = tensor([1, 1, 8, 128])]; + tensor v_117 = reshape(shape = var_7947, x = v_115)[name = string("v_117")]; + fp32 var_7952_promoted = const()[name = string("op_7952_promoted"), val = fp32(0x1p+1)]; + tensor var_7958 = pow(x = hidden_states_463, y = var_7952_promoted)[name = string("op_7958")]; + tensor variance_155_axes_0 = const()[name = string("variance_155_axes_0"), val = tensor([-1])]; + bool variance_155_keep_dims_0 = const()[name = string("variance_155_keep_dims_0"), val = bool(true)]; + tensor variance_155 = reduce_mean(axes = variance_155_axes_0, keep_dims = variance_155_keep_dims_0, x = var_7958)[name = string("variance_155")]; + fp32 var_7961 = const()[name = string("op_7961"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7962 = add(x = variance_155, y = var_7961)[name = string("op_7962")]; + fp32 var_7963_epsilon_0 = const()[name = string("op_7963_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7963 = rsqrt(epsilon = var_7963_epsilon_0, x = var_7962)[name = string("op_7963")]; + tensor hidden_states_467 = mul(x = hidden_states_463, y = var_7963)[name = string("hidden_states_467")]; + tensor const_192 = const()[name = string("const_192"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774412480)))]; + tensor q_155 = mul(x = const_192, y = hidden_states_467)[name = string("q_155")]; + fp32 var_7970_promoted = const()[name = string("op_7970_promoted"), val = fp32(0x1p+1)]; + tensor var_7976 = pow(x = hidden_states_469, y = var_7970_promoted)[name = string("op_7976")]; + tensor variance_157_axes_0 = const()[name = string("variance_157_axes_0"), val = tensor([-1])]; + bool variance_157_keep_dims_0 = const()[name = string("variance_157_keep_dims_0"), val = bool(true)]; + tensor variance_157 = reduce_mean(axes = variance_157_axes_0, keep_dims = variance_157_keep_dims_0, x = var_7976)[name = string("variance_157")]; + fp32 var_7979 = const()[name = string("op_7979"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_7980 = add(x = variance_157, y = var_7979)[name = string("op_7980")]; + fp32 var_7981_epsilon_0 = const()[name = string("op_7981_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_7981 = rsqrt(epsilon = var_7981_epsilon_0, x = var_7980)[name = string("op_7981")]; + tensor hidden_states_473 = mul(x = hidden_states_469, y = var_7981)[name = string("hidden_states_473")]; + tensor const_193 = const()[name = string("const_193"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774413056)))]; + tensor k_155 = mul(x = const_193, y = hidden_states_473)[name = string("k_155")]; + tensor q_157_perm_0 = const()[name = string("q_157_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_157_perm_0 = const()[name = string("k_157_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_119_perm_0 = const()[name = string("v_119_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_157 = transpose(perm = q_157_perm_0, x = q_155)[name = string("transpose_35")]; + tensor var_7998 = mul(x = q_157, y = cos_3)[name = string("op_7998")]; + tensor x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_77 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_157)[name = string("x1_77")]; + tensor x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_77 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_157)[name = string("x2_77")]; + fp32 const_196_promoted = const()[name = string("const_196_promoted"), val = fp32(-0x1p+0)]; + tensor var_8019 = mul(x = x2_77, y = const_196_promoted)[name = string("op_8019")]; + int32 var_8021 = const()[name = string("op_8021"), val = int32(-1)]; + bool var_8022_interleave_0 = const()[name = string("op_8022_interleave_0"), val = bool(false)]; + tensor var_8022 = concat(axis = var_8021, interleave = var_8022_interleave_0, values = (var_8019, x1_77))[name = string("op_8022")]; + tensor var_8023 = mul(x = var_8022, y = sin_3)[name = string("op_8023")]; + tensor q_159 = add(x = var_7998, y = var_8023)[name = string("q_159")]; + tensor k_157 = transpose(perm = k_157_perm_0, x = k_155)[name = string("transpose_34")]; + tensor var_8026 = mul(x = k_157, y = cos_3)[name = string("op_8026")]; + tensor x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_79 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_157)[name = string("x1_79")]; + tensor x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_79 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_157)[name = string("x2_79")]; + fp32 const_199_promoted = const()[name = string("const_199_promoted"), val = fp32(-0x1p+0)]; + tensor var_8047 = mul(x = x2_79, y = const_199_promoted)[name = string("op_8047")]; + int32 var_8049 = const()[name = string("op_8049"), val = int32(-1)]; + bool var_8050_interleave_0 = const()[name = string("op_8050_interleave_0"), val = bool(false)]; + tensor var_8050 = concat(axis = var_8049, interleave = var_8050_interleave_0, values = (var_8047, x1_79))[name = string("op_8050")]; + tensor var_8051 = mul(x = var_8050, y = sin_3)[name = string("op_8051")]; + tensor k_159 = add(x = var_8026, y = var_8051)[name = string("k_159")]; + tensor var_8058 = const()[name = string("op_8058"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_39 = reshape(shape = var_8058, x = k_159)[name = string("nk_flat_39")]; + tensor var_8064 = const()[name = string("op_8064"), val = tensor([1, 1024, 1, 1])]; + tensor v_119 = transpose(perm = v_119_perm_0, x = v_117)[name = string("transpose_33")]; + tensor nv_flat_39 = reshape(shape = var_8064, x = v_119)[name = string("nv_flat_39")]; + tensor var_8073 = mul(x = var_7867, y = var_1194)[name = string("op_8073")]; + tensor var_8074 = mul(x = nk_flat_39, y = update_mask_1)[name = string("op_8074")]; + tensor key_cache_81 = add(x = var_8073, y = var_8074)[name = string("key_cache_81")]; + tensor var_8080 = mul(x = var_7887, y = var_1194)[name = string("op_8080")]; + tensor var_8081 = mul(x = nv_flat_39, y = update_mask_1)[name = string("op_8081")]; + tensor value_cache_81 = add(x = var_8080, y = var_8081)[name = string("value_cache_81")]; + tensor kc_115_axes_0 = const()[name = string("kc_115_axes_0"), val = tensor([2])]; + tensor kc_115 = squeeze(axes = kc_115_axes_0, x = key_cache_81)[name = string("kc_115")]; + tensor var_8090 = const()[name = string("op_8090"), val = tensor([1, 8, 128, 256])]; + tensor kc_117 = reshape(shape = var_8090, x = kc_115)[name = string("kc_117")]; + tensor vc_115_axes_0 = const()[name = string("vc_115_axes_0"), val = tensor([2])]; + tensor vc_115 = squeeze(axes = vc_115_axes_0, x = value_cache_81)[name = string("vc_115")]; + tensor var_8098 = const()[name = string("op_8098"), val = tensor([1, 8, 128, 256])]; + tensor vc_117 = reshape(shape = var_8098, x = vc_115)[name = string("vc_117")]; + tensor var_8101_axes_0 = const()[name = string("op_8101_axes_0"), val = tensor([2])]; + tensor var_8101 = expand_dims(axes = var_8101_axes_0, x = kc_117)[name = string("op_8101")]; + tensor var_8109_reps_0 = const()[name = string("op_8109_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_8109 = tile(reps = var_8109_reps_0, x = var_8101)[name = string("op_8109")]; + tensor var_8114 = const()[name = string("op_8114"), val = tensor([1, 16, 128, 256])]; + tensor kc_119 = reshape(shape = var_8114, x = var_8109)[name = string("kc_119")]; + tensor var_8117_axes_0 = const()[name = string("op_8117_axes_0"), val = tensor([2])]; + tensor var_8117 = expand_dims(axes = var_8117_axes_0, x = vc_117)[name = string("op_8117")]; + tensor var_8125_reps_0 = const()[name = string("op_8125_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_8125 = tile(reps = var_8125_reps_0, x = var_8117)[name = string("op_8125")]; + tensor var_8130 = const()[name = string("op_8130"), val = tensor([1, 16, 128, 256])]; + tensor vc_119 = reshape(shape = var_8130, x = var_8125)[name = string("vc_119")]; + bool var_8132_transpose_x_0 = const()[name = string("op_8132_transpose_x_0"), val = bool(false)]; + bool var_8132_transpose_y_0 = const()[name = string("op_8132_transpose_y_0"), val = bool(false)]; + tensor var_8132 = matmul(transpose_x = var_8132_transpose_x_0, transpose_y = var_8132_transpose_y_0, x = q_159, y = kc_119)[name = string("op_8132")]; + fp32 _inversed_attn_weights_153_y_0 = const()[name = string("_inversed_attn_weights_153_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_153 = mul(x = var_8132, y = _inversed_attn_weights_153_y_0)[name = string("_inversed_attn_weights_153")]; + tensor attn_weights_155 = add(x = _inversed_attn_weights_153, y = mask_1)[name = string("attn_weights_155")]; + int32 var_8146 = const()[name = string("op_8146"), val = int32(-1)]; + tensor attn_weights_159 = softmax(axis = var_8146, x = attn_weights_155)[name = string("attn_weights_159")]; + bool attn_output_77_transpose_x_1 = const()[name = string("attn_output_77_transpose_x_1"), val = bool(false)]; + bool attn_output_77_transpose_y_1 = const()[name = string("attn_output_77_transpose_y_1"), val = bool(true)]; + tensor attn_output_77 = matmul(transpose_x = attn_output_77_transpose_x_1, transpose_y = attn_output_77_transpose_y_1, x = attn_weights_159, y = vc_119)[name = string("attn_output_77")]; + tensor var_8155_perm_0 = const()[name = string("op_8155_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_8159 = const()[name = string("op_8159"), val = tensor([1, 1, -1])]; + tensor var_8155 = transpose(perm = var_8155_perm_0, x = attn_output_77)[name = string("transpose_32")]; + tensor input_193 = reshape(shape = var_8159, x = var_8155)[name = string("input_193")]; + tensor attn_output_79 = linear(bias = linear_1_bias_0, weight = layers_19_self_attn_o_proj_weight, x = input_193)[name = string("linear_136")]; + tensor var_8165_axes_0 = const()[name = string("op_8165_axes_0"), val = tensor([0])]; + tensor var_8165 = squeeze(axes = var_8165_axes_0, x = attn_output_79)[name = string("op_8165")]; + tensor var_8167_axes_0 = const()[name = string("op_8167_axes_0"), val = tensor([0])]; + tensor var_8167 = squeeze(axes = var_8167_axes_0, x = var_8165)[name = string("op_8167")]; + tensor var_8169_axes_0 = const()[name = string("op_8169_axes_0"), val = tensor([-1])]; + tensor var_8169 = expand_dims(axes = var_8169_axes_0, x = var_8167)[name = string("op_8169")]; + tensor attn_4d_39_axes_0 = const()[name = string("attn_4d_39_axes_0"), val = tensor([-1])]; + tensor attn_4d_39 = expand_dims(axes = attn_4d_39_axes_0, x = var_8169)[name = string("attn_4d_39")]; + tensor hidden_77 = add(x = hidden_75, y = attn_4d_39)[name = string("hidden_77")]; + tensor var_8175_axes_0 = const()[name = string("op_8175_axes_0"), val = tensor([-1])]; + tensor var_8175 = squeeze(axes = var_8175_axes_0, x = hidden_77)[name = string("op_8175")]; + tensor var_8177_axes_0 = const()[name = string("op_8177_axes_0"), val = tensor([-1])]; + tensor var_8177 = squeeze(axes = var_8177_axes_0, x = var_8175)[name = string("op_8177")]; + tensor hidden_states_475_axes_0 = const()[name = string("hidden_states_475_axes_0"), val = tensor([0])]; + tensor hidden_states_475 = expand_dims(axes = hidden_states_475_axes_0, x = var_8177)[name = string("hidden_states_475")]; + fp32 var_8183_promoted = const()[name = string("op_8183_promoted"), val = fp32(0x1p+1)]; + tensor var_8189 = pow(x = hidden_states_475, y = var_8183_promoted)[name = string("op_8189")]; + tensor variance_159_axes_0 = const()[name = string("variance_159_axes_0"), val = tensor([-1])]; + bool variance_159_keep_dims_0 = const()[name = string("variance_159_keep_dims_0"), val = bool(true)]; + tensor variance_159 = reduce_mean(axes = variance_159_axes_0, keep_dims = variance_159_keep_dims_0, x = var_8189)[name = string("variance_159")]; + fp32 var_8192 = const()[name = string("op_8192"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8193 = add(x = variance_159, y = var_8192)[name = string("op_8193")]; + fp32 var_8194_epsilon_0 = const()[name = string("op_8194_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8194 = rsqrt(epsilon = var_8194_epsilon_0, x = var_8193)[name = string("op_8194")]; + tensor hidden_states_479 = mul(x = hidden_states_475, y = var_8194)[name = string("hidden_states_479")]; + tensor const_200 = const()[name = string("const_200"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774413632)))]; + tensor input_195 = mul(x = const_200, y = hidden_states_479)[name = string("input_195")]; + tensor input_197 = linear(bias = linear_4_bias_0, weight = layers_19_mlp_gate_proj_weight, x = input_195)[name = string("linear_137")]; + tensor var_8204 = silu(x = input_197)[name = string("op_8204")]; + tensor var_8206 = linear(bias = linear_4_bias_0, weight = layers_19_mlp_up_proj_weight, x = input_195)[name = string("linear_138")]; + tensor input_199 = mul(x = var_8204, y = var_8206)[name = string("input_199")]; + tensor mlp_out_39 = linear(bias = linear_1_bias_0, weight = layers_19_mlp_down_proj_weight, x = input_199)[name = string("linear_139")]; + tensor var_8211_axes_0 = const()[name = string("op_8211_axes_0"), val = tensor([0])]; + tensor var_8211 = squeeze(axes = var_8211_axes_0, x = mlp_out_39)[name = string("op_8211")]; + tensor var_8213_axes_0 = const()[name = string("op_8213_axes_0"), val = tensor([0])]; + tensor var_8213 = squeeze(axes = var_8213_axes_0, x = var_8211)[name = string("op_8213")]; + tensor var_8215_axes_0 = const()[name = string("op_8215_axes_0"), val = tensor([-1])]; + tensor var_8215 = expand_dims(axes = var_8215_axes_0, x = var_8213)[name = string("op_8215")]; + tensor mlp_4d_39_axes_0 = const()[name = string("mlp_4d_39_axes_0"), val = tensor([-1])]; + tensor mlp_4d_39 = expand_dims(axes = mlp_4d_39_axes_0, x = var_8215)[name = string("mlp_4d_39")]; + tensor hidden_79 = add(x = hidden_77, y = mlp_4d_39)[name = string("hidden_79")]; + tensor var_8229_begin_0 = const()[name = string("op_8229_begin_0"), val = tensor([0, 20480, 0, 0])]; + tensor var_8229_end_0 = const()[name = string("op_8229_end_0"), val = tensor([1, 21504, 1, 256])]; + tensor var_8229_end_mask_0 = const()[name = string("op_8229_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_8229 = slice_by_index(begin = var_8229_begin_0, end = var_8229_end_0, end_mask = var_8229_end_mask_0, x = cast_3)[name = string("op_8229")]; + tensor var_8249_begin_0 = const()[name = string("op_8249_begin_0"), val = tensor([0, 20480, 0, 0])]; + tensor var_8249_end_0 = const()[name = string("op_8249_end_0"), val = tensor([1, 21504, 1, 256])]; + tensor var_8249_end_mask_0 = const()[name = string("op_8249_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_8249 = slice_by_index(begin = var_8249_begin_0, end = var_8249_end_0, end_mask = var_8249_end_mask_0, x = cast_4)[name = string("op_8249")]; + tensor var_8261_axes_0 = const()[name = string("op_8261_axes_0"), val = tensor([-1])]; + tensor var_8261 = squeeze(axes = var_8261_axes_0, x = hidden_79)[name = string("op_8261")]; + tensor var_8263_axes_0 = const()[name = string("op_8263_axes_0"), val = tensor([-1])]; + tensor var_8263 = squeeze(axes = var_8263_axes_0, x = var_8261)[name = string("op_8263")]; + tensor hidden_states_481_axes_0 = const()[name = string("hidden_states_481_axes_0"), val = tensor([0])]; + tensor hidden_states_481 = expand_dims(axes = hidden_states_481_axes_0, x = var_8263)[name = string("hidden_states_481")]; + fp32 var_8269_promoted = const()[name = string("op_8269_promoted"), val = fp32(0x1p+1)]; + tensor var_8275 = pow(x = hidden_states_481, y = var_8269_promoted)[name = string("op_8275")]; + tensor variance_161_axes_0 = const()[name = string("variance_161_axes_0"), val = tensor([-1])]; + bool variance_161_keep_dims_0 = const()[name = string("variance_161_keep_dims_0"), val = bool(true)]; + tensor variance_161 = reduce_mean(axes = variance_161_axes_0, keep_dims = variance_161_keep_dims_0, x = var_8275)[name = string("variance_161")]; + fp32 var_8278 = const()[name = string("op_8278"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8279 = add(x = variance_161, y = var_8278)[name = string("op_8279")]; + fp32 var_8280_epsilon_0 = const()[name = string("op_8280_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8280 = rsqrt(epsilon = var_8280_epsilon_0, x = var_8279)[name = string("op_8280")]; + tensor hidden_states_485 = mul(x = hidden_states_481, y = var_8280)[name = string("hidden_states_485")]; + tensor const_201 = const()[name = string("const_201"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774417792)))]; + tensor input_201 = mul(x = const_201, y = hidden_states_485)[name = string("input_201")]; + tensor q_161 = linear(bias = linear_0_bias_0, weight = layers_20_self_attn_q_proj_weight, x = input_201)[name = string("linear_140")]; + tensor k_161 = linear(bias = linear_1_bias_0, weight = layers_20_self_attn_k_proj_weight, x = input_201)[name = string("linear_141")]; + tensor v_121 = linear(bias = linear_1_bias_0, weight = layers_20_self_attn_v_proj_weight, x = input_201)[name = string("linear_142")]; + tensor var_8297 = const()[name = string("op_8297"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_487 = reshape(shape = var_8297, x = q_161)[name = string("hidden_states_487")]; + tensor var_8303 = const()[name = string("op_8303"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_493 = reshape(shape = var_8303, x = k_161)[name = string("hidden_states_493")]; + tensor var_8309 = const()[name = string("op_8309"), val = tensor([1, 1, 8, 128])]; + tensor v_123 = reshape(shape = var_8309, x = v_121)[name = string("v_123")]; + fp32 var_8314_promoted = const()[name = string("op_8314_promoted"), val = fp32(0x1p+1)]; + tensor var_8320 = pow(x = hidden_states_487, y = var_8314_promoted)[name = string("op_8320")]; + tensor variance_163_axes_0 = const()[name = string("variance_163_axes_0"), val = tensor([-1])]; + bool variance_163_keep_dims_0 = const()[name = string("variance_163_keep_dims_0"), val = bool(true)]; + tensor variance_163 = reduce_mean(axes = variance_163_axes_0, keep_dims = variance_163_keep_dims_0, x = var_8320)[name = string("variance_163")]; + fp32 var_8323 = const()[name = string("op_8323"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8324 = add(x = variance_163, y = var_8323)[name = string("op_8324")]; + fp32 var_8325_epsilon_0 = const()[name = string("op_8325_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8325 = rsqrt(epsilon = var_8325_epsilon_0, x = var_8324)[name = string("op_8325")]; + tensor hidden_states_491 = mul(x = hidden_states_487, y = var_8325)[name = string("hidden_states_491")]; + tensor const_202 = const()[name = string("const_202"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774421952)))]; + tensor q_163 = mul(x = const_202, y = hidden_states_491)[name = string("q_163")]; + fp32 var_8332_promoted = const()[name = string("op_8332_promoted"), val = fp32(0x1p+1)]; + tensor var_8338 = pow(x = hidden_states_493, y = var_8332_promoted)[name = string("op_8338")]; + tensor variance_165_axes_0 = const()[name = string("variance_165_axes_0"), val = tensor([-1])]; + bool variance_165_keep_dims_0 = const()[name = string("variance_165_keep_dims_0"), val = bool(true)]; + tensor variance_165 = reduce_mean(axes = variance_165_axes_0, keep_dims = variance_165_keep_dims_0, x = var_8338)[name = string("variance_165")]; + fp32 var_8341 = const()[name = string("op_8341"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8342 = add(x = variance_165, y = var_8341)[name = string("op_8342")]; + fp32 var_8343_epsilon_0 = const()[name = string("op_8343_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8343 = rsqrt(epsilon = var_8343_epsilon_0, x = var_8342)[name = string("op_8343")]; + tensor hidden_states_497 = mul(x = hidden_states_493, y = var_8343)[name = string("hidden_states_497")]; + tensor const_203 = const()[name = string("const_203"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774422528)))]; + tensor k_163 = mul(x = const_203, y = hidden_states_497)[name = string("k_163")]; + tensor q_165_perm_0 = const()[name = string("q_165_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_165_perm_0 = const()[name = string("k_165_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_125_perm_0 = const()[name = string("v_125_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_165 = transpose(perm = q_165_perm_0, x = q_163)[name = string("transpose_31")]; + tensor var_8360 = mul(x = q_165, y = cos_3)[name = string("op_8360")]; + tensor x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_81 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_165)[name = string("x1_81")]; + tensor x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_81 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_165)[name = string("x2_81")]; + fp32 const_206_promoted = const()[name = string("const_206_promoted"), val = fp32(-0x1p+0)]; + tensor var_8381 = mul(x = x2_81, y = const_206_promoted)[name = string("op_8381")]; + int32 var_8383 = const()[name = string("op_8383"), val = int32(-1)]; + bool var_8384_interleave_0 = const()[name = string("op_8384_interleave_0"), val = bool(false)]; + tensor var_8384 = concat(axis = var_8383, interleave = var_8384_interleave_0, values = (var_8381, x1_81))[name = string("op_8384")]; + tensor var_8385 = mul(x = var_8384, y = sin_3)[name = string("op_8385")]; + tensor q_167 = add(x = var_8360, y = var_8385)[name = string("q_167")]; + tensor k_165 = transpose(perm = k_165_perm_0, x = k_163)[name = string("transpose_30")]; + tensor var_8388 = mul(x = k_165, y = cos_3)[name = string("op_8388")]; + tensor x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_83 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_165)[name = string("x1_83")]; + tensor x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_83 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_165)[name = string("x2_83")]; + fp32 const_209_promoted = const()[name = string("const_209_promoted"), val = fp32(-0x1p+0)]; + tensor var_8409 = mul(x = x2_83, y = const_209_promoted)[name = string("op_8409")]; + int32 var_8411 = const()[name = string("op_8411"), val = int32(-1)]; + bool var_8412_interleave_0 = const()[name = string("op_8412_interleave_0"), val = bool(false)]; + tensor var_8412 = concat(axis = var_8411, interleave = var_8412_interleave_0, values = (var_8409, x1_83))[name = string("op_8412")]; + tensor var_8413 = mul(x = var_8412, y = sin_3)[name = string("op_8413")]; + tensor k_167 = add(x = var_8388, y = var_8413)[name = string("k_167")]; + tensor var_8420 = const()[name = string("op_8420"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_41 = reshape(shape = var_8420, x = k_167)[name = string("nk_flat_41")]; + tensor var_8426 = const()[name = string("op_8426"), val = tensor([1, 1024, 1, 1])]; + tensor v_125 = transpose(perm = v_125_perm_0, x = v_123)[name = string("transpose_29")]; + tensor nv_flat_41 = reshape(shape = var_8426, x = v_125)[name = string("nv_flat_41")]; + tensor var_8435 = mul(x = var_8229, y = var_1194)[name = string("op_8435")]; + tensor var_8436 = mul(x = nk_flat_41, y = update_mask_1)[name = string("op_8436")]; + tensor key_cache_85 = add(x = var_8435, y = var_8436)[name = string("key_cache_85")]; + tensor var_8442 = mul(x = var_8249, y = var_1194)[name = string("op_8442")]; + tensor var_8443 = mul(x = nv_flat_41, y = update_mask_1)[name = string("op_8443")]; + tensor value_cache_85 = add(x = var_8442, y = var_8443)[name = string("value_cache_85")]; + tensor kc_121_axes_0 = const()[name = string("kc_121_axes_0"), val = tensor([2])]; + tensor kc_121 = squeeze(axes = kc_121_axes_0, x = key_cache_85)[name = string("kc_121")]; + tensor var_8452 = const()[name = string("op_8452"), val = tensor([1, 8, 128, 256])]; + tensor kc_123 = reshape(shape = var_8452, x = kc_121)[name = string("kc_123")]; + tensor vc_121_axes_0 = const()[name = string("vc_121_axes_0"), val = tensor([2])]; + tensor vc_121 = squeeze(axes = vc_121_axes_0, x = value_cache_85)[name = string("vc_121")]; + tensor var_8460 = const()[name = string("op_8460"), val = tensor([1, 8, 128, 256])]; + tensor vc_123 = reshape(shape = var_8460, x = vc_121)[name = string("vc_123")]; + tensor var_8463_axes_0 = const()[name = string("op_8463_axes_0"), val = tensor([2])]; + tensor var_8463 = expand_dims(axes = var_8463_axes_0, x = kc_123)[name = string("op_8463")]; + tensor var_8471_reps_0 = const()[name = string("op_8471_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_8471 = tile(reps = var_8471_reps_0, x = var_8463)[name = string("op_8471")]; + tensor var_8476 = const()[name = string("op_8476"), val = tensor([1, 16, 128, 256])]; + tensor kc_125 = reshape(shape = var_8476, x = var_8471)[name = string("kc_125")]; + tensor var_8479_axes_0 = const()[name = string("op_8479_axes_0"), val = tensor([2])]; + tensor var_8479 = expand_dims(axes = var_8479_axes_0, x = vc_123)[name = string("op_8479")]; + tensor var_8487_reps_0 = const()[name = string("op_8487_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_8487 = tile(reps = var_8487_reps_0, x = var_8479)[name = string("op_8487")]; + tensor var_8492 = const()[name = string("op_8492"), val = tensor([1, 16, 128, 256])]; + tensor vc_125 = reshape(shape = var_8492, x = var_8487)[name = string("vc_125")]; + bool var_8494_transpose_x_0 = const()[name = string("op_8494_transpose_x_0"), val = bool(false)]; + bool var_8494_transpose_y_0 = const()[name = string("op_8494_transpose_y_0"), val = bool(false)]; + tensor var_8494 = matmul(transpose_x = var_8494_transpose_x_0, transpose_y = var_8494_transpose_y_0, x = q_167, y = kc_125)[name = string("op_8494")]; + fp32 _inversed_attn_weights_161_y_0 = const()[name = string("_inversed_attn_weights_161_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_161 = mul(x = var_8494, y = _inversed_attn_weights_161_y_0)[name = string("_inversed_attn_weights_161")]; + tensor attn_weights_163 = add(x = _inversed_attn_weights_161, y = mask_1)[name = string("attn_weights_163")]; + int32 var_8508 = const()[name = string("op_8508"), val = int32(-1)]; + tensor attn_weights_167 = softmax(axis = var_8508, x = attn_weights_163)[name = string("attn_weights_167")]; + bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(false)]; + bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(true)]; + tensor attn_output_81 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_167, y = vc_125)[name = string("attn_output_81")]; + tensor var_8517_perm_0 = const()[name = string("op_8517_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_8521 = const()[name = string("op_8521"), val = tensor([1, 1, -1])]; + tensor var_8517 = transpose(perm = var_8517_perm_0, x = attn_output_81)[name = string("transpose_28")]; + tensor input_203 = reshape(shape = var_8521, x = var_8517)[name = string("input_203")]; + tensor attn_output_83 = linear(bias = linear_1_bias_0, weight = layers_20_self_attn_o_proj_weight, x = input_203)[name = string("linear_143")]; + tensor var_8527_axes_0 = const()[name = string("op_8527_axes_0"), val = tensor([0])]; + tensor var_8527 = squeeze(axes = var_8527_axes_0, x = attn_output_83)[name = string("op_8527")]; + tensor var_8529_axes_0 = const()[name = string("op_8529_axes_0"), val = tensor([0])]; + tensor var_8529 = squeeze(axes = var_8529_axes_0, x = var_8527)[name = string("op_8529")]; + tensor var_8531_axes_0 = const()[name = string("op_8531_axes_0"), val = tensor([-1])]; + tensor var_8531 = expand_dims(axes = var_8531_axes_0, x = var_8529)[name = string("op_8531")]; + tensor attn_4d_41_axes_0 = const()[name = string("attn_4d_41_axes_0"), val = tensor([-1])]; + tensor attn_4d_41 = expand_dims(axes = attn_4d_41_axes_0, x = var_8531)[name = string("attn_4d_41")]; + tensor hidden_81 = add(x = hidden_79, y = attn_4d_41)[name = string("hidden_81")]; + tensor var_8537_axes_0 = const()[name = string("op_8537_axes_0"), val = tensor([-1])]; + tensor var_8537 = squeeze(axes = var_8537_axes_0, x = hidden_81)[name = string("op_8537")]; + tensor var_8539_axes_0 = const()[name = string("op_8539_axes_0"), val = tensor([-1])]; + tensor var_8539 = squeeze(axes = var_8539_axes_0, x = var_8537)[name = string("op_8539")]; + tensor hidden_states_499_axes_0 = const()[name = string("hidden_states_499_axes_0"), val = tensor([0])]; + tensor hidden_states_499 = expand_dims(axes = hidden_states_499_axes_0, x = var_8539)[name = string("hidden_states_499")]; + fp32 var_8545_promoted = const()[name = string("op_8545_promoted"), val = fp32(0x1p+1)]; + tensor var_8551 = pow(x = hidden_states_499, y = var_8545_promoted)[name = string("op_8551")]; + tensor variance_167_axes_0 = const()[name = string("variance_167_axes_0"), val = tensor([-1])]; + bool variance_167_keep_dims_0 = const()[name = string("variance_167_keep_dims_0"), val = bool(true)]; + tensor variance_167 = reduce_mean(axes = variance_167_axes_0, keep_dims = variance_167_keep_dims_0, x = var_8551)[name = string("variance_167")]; + fp32 var_8554 = const()[name = string("op_8554"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8555 = add(x = variance_167, y = var_8554)[name = string("op_8555")]; + fp32 var_8556_epsilon_0 = const()[name = string("op_8556_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8556 = rsqrt(epsilon = var_8556_epsilon_0, x = var_8555)[name = string("op_8556")]; + tensor hidden_states_503 = mul(x = hidden_states_499, y = var_8556)[name = string("hidden_states_503")]; + tensor const_210 = const()[name = string("const_210"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774423104)))]; + tensor input_205 = mul(x = const_210, y = hidden_states_503)[name = string("input_205")]; + tensor input_207 = linear(bias = linear_4_bias_0, weight = layers_20_mlp_gate_proj_weight, x = input_205)[name = string("linear_144")]; + tensor var_8566 = silu(x = input_207)[name = string("op_8566")]; + tensor var_8568 = linear(bias = linear_4_bias_0, weight = layers_20_mlp_up_proj_weight, x = input_205)[name = string("linear_145")]; + tensor input_209 = mul(x = var_8566, y = var_8568)[name = string("input_209")]; + tensor mlp_out_41 = linear(bias = linear_1_bias_0, weight = layers_20_mlp_down_proj_weight, x = input_209)[name = string("linear_146")]; + tensor var_8573_axes_0 = const()[name = string("op_8573_axes_0"), val = tensor([0])]; + tensor var_8573 = squeeze(axes = var_8573_axes_0, x = mlp_out_41)[name = string("op_8573")]; + tensor var_8575_axes_0 = const()[name = string("op_8575_axes_0"), val = tensor([0])]; + tensor var_8575 = squeeze(axes = var_8575_axes_0, x = var_8573)[name = string("op_8575")]; + tensor var_8577_axes_0 = const()[name = string("op_8577_axes_0"), val = tensor([-1])]; + tensor var_8577 = expand_dims(axes = var_8577_axes_0, x = var_8575)[name = string("op_8577")]; + tensor mlp_4d_41_axes_0 = const()[name = string("mlp_4d_41_axes_0"), val = tensor([-1])]; + tensor mlp_4d_41 = expand_dims(axes = mlp_4d_41_axes_0, x = var_8577)[name = string("mlp_4d_41")]; + tensor hidden_83 = add(x = hidden_81, y = mlp_4d_41)[name = string("hidden_83")]; + tensor var_8591_begin_0 = const()[name = string("op_8591_begin_0"), val = tensor([0, 21504, 0, 0])]; + tensor var_8591_end_0 = const()[name = string("op_8591_end_0"), val = tensor([1, 22528, 1, 256])]; + tensor var_8591_end_mask_0 = const()[name = string("op_8591_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_8591 = slice_by_index(begin = var_8591_begin_0, end = var_8591_end_0, end_mask = var_8591_end_mask_0, x = cast_3)[name = string("op_8591")]; + tensor var_8611_begin_0 = const()[name = string("op_8611_begin_0"), val = tensor([0, 21504, 0, 0])]; + tensor var_8611_end_0 = const()[name = string("op_8611_end_0"), val = tensor([1, 22528, 1, 256])]; + tensor var_8611_end_mask_0 = const()[name = string("op_8611_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_8611 = slice_by_index(begin = var_8611_begin_0, end = var_8611_end_0, end_mask = var_8611_end_mask_0, x = cast_4)[name = string("op_8611")]; + tensor var_8623_axes_0 = const()[name = string("op_8623_axes_0"), val = tensor([-1])]; + tensor var_8623 = squeeze(axes = var_8623_axes_0, x = hidden_83)[name = string("op_8623")]; + tensor var_8625_axes_0 = const()[name = string("op_8625_axes_0"), val = tensor([-1])]; + tensor var_8625 = squeeze(axes = var_8625_axes_0, x = var_8623)[name = string("op_8625")]; + tensor hidden_states_505_axes_0 = const()[name = string("hidden_states_505_axes_0"), val = tensor([0])]; + tensor hidden_states_505 = expand_dims(axes = hidden_states_505_axes_0, x = var_8625)[name = string("hidden_states_505")]; + fp32 var_8631_promoted = const()[name = string("op_8631_promoted"), val = fp32(0x1p+1)]; + tensor var_8637 = pow(x = hidden_states_505, y = var_8631_promoted)[name = string("op_8637")]; + tensor variance_169_axes_0 = const()[name = string("variance_169_axes_0"), val = tensor([-1])]; + bool variance_169_keep_dims_0 = const()[name = string("variance_169_keep_dims_0"), val = bool(true)]; + tensor variance_169 = reduce_mean(axes = variance_169_axes_0, keep_dims = variance_169_keep_dims_0, x = var_8637)[name = string("variance_169")]; + fp32 var_8640 = const()[name = string("op_8640"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8641 = add(x = variance_169, y = var_8640)[name = string("op_8641")]; + fp32 var_8642_epsilon_0 = const()[name = string("op_8642_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8642 = rsqrt(epsilon = var_8642_epsilon_0, x = var_8641)[name = string("op_8642")]; + tensor hidden_states_509 = mul(x = hidden_states_505, y = var_8642)[name = string("hidden_states_509")]; + tensor const_211 = const()[name = string("const_211"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774427264)))]; + tensor input_211 = mul(x = const_211, y = hidden_states_509)[name = string("input_211")]; + tensor q_169 = linear(bias = linear_0_bias_0, weight = layers_21_self_attn_q_proj_weight, x = input_211)[name = string("linear_147")]; + tensor k_169 = linear(bias = linear_1_bias_0, weight = layers_21_self_attn_k_proj_weight, x = input_211)[name = string("linear_148")]; + tensor v_127 = linear(bias = linear_1_bias_0, weight = layers_21_self_attn_v_proj_weight, x = input_211)[name = string("linear_149")]; + tensor var_8659 = const()[name = string("op_8659"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_511 = reshape(shape = var_8659, x = q_169)[name = string("hidden_states_511")]; + tensor var_8665 = const()[name = string("op_8665"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_517 = reshape(shape = var_8665, x = k_169)[name = string("hidden_states_517")]; + tensor var_8671 = const()[name = string("op_8671"), val = tensor([1, 1, 8, 128])]; + tensor v_129 = reshape(shape = var_8671, x = v_127)[name = string("v_129")]; + fp32 var_8676_promoted = const()[name = string("op_8676_promoted"), val = fp32(0x1p+1)]; + tensor var_8682 = pow(x = hidden_states_511, y = var_8676_promoted)[name = string("op_8682")]; + tensor variance_171_axes_0 = const()[name = string("variance_171_axes_0"), val = tensor([-1])]; + bool variance_171_keep_dims_0 = const()[name = string("variance_171_keep_dims_0"), val = bool(true)]; + tensor variance_171 = reduce_mean(axes = variance_171_axes_0, keep_dims = variance_171_keep_dims_0, x = var_8682)[name = string("variance_171")]; + fp32 var_8685 = const()[name = string("op_8685"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8686 = add(x = variance_171, y = var_8685)[name = string("op_8686")]; + fp32 var_8687_epsilon_0 = const()[name = string("op_8687_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8687 = rsqrt(epsilon = var_8687_epsilon_0, x = var_8686)[name = string("op_8687")]; + tensor hidden_states_515 = mul(x = hidden_states_511, y = var_8687)[name = string("hidden_states_515")]; + tensor const_212 = const()[name = string("const_212"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774431424)))]; + tensor q_171 = mul(x = const_212, y = hidden_states_515)[name = string("q_171")]; + fp32 var_8694_promoted = const()[name = string("op_8694_promoted"), val = fp32(0x1p+1)]; + tensor var_8700 = pow(x = hidden_states_517, y = var_8694_promoted)[name = string("op_8700")]; + tensor variance_173_axes_0 = const()[name = string("variance_173_axes_0"), val = tensor([-1])]; + bool variance_173_keep_dims_0 = const()[name = string("variance_173_keep_dims_0"), val = bool(true)]; + tensor variance_173 = reduce_mean(axes = variance_173_axes_0, keep_dims = variance_173_keep_dims_0, x = var_8700)[name = string("variance_173")]; + fp32 var_8703 = const()[name = string("op_8703"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8704 = add(x = variance_173, y = var_8703)[name = string("op_8704")]; + fp32 var_8705_epsilon_0 = const()[name = string("op_8705_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8705 = rsqrt(epsilon = var_8705_epsilon_0, x = var_8704)[name = string("op_8705")]; + tensor hidden_states_521 = mul(x = hidden_states_517, y = var_8705)[name = string("hidden_states_521")]; + tensor const_213 = const()[name = string("const_213"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774432000)))]; + tensor k_171 = mul(x = const_213, y = hidden_states_521)[name = string("k_171")]; + tensor q_173_perm_0 = const()[name = string("q_173_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_173_perm_0 = const()[name = string("k_173_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_131_perm_0 = const()[name = string("v_131_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_173 = transpose(perm = q_173_perm_0, x = q_171)[name = string("transpose_27")]; + tensor var_8722 = mul(x = q_173, y = cos_3)[name = string("op_8722")]; + tensor x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_85 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_173)[name = string("x1_85")]; + tensor x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_85 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_173)[name = string("x2_85")]; + fp32 const_216_promoted = const()[name = string("const_216_promoted"), val = fp32(-0x1p+0)]; + tensor var_8743 = mul(x = x2_85, y = const_216_promoted)[name = string("op_8743")]; + int32 var_8745 = const()[name = string("op_8745"), val = int32(-1)]; + bool var_8746_interleave_0 = const()[name = string("op_8746_interleave_0"), val = bool(false)]; + tensor var_8746 = concat(axis = var_8745, interleave = var_8746_interleave_0, values = (var_8743, x1_85))[name = string("op_8746")]; + tensor var_8747 = mul(x = var_8746, y = sin_3)[name = string("op_8747")]; + tensor q_175 = add(x = var_8722, y = var_8747)[name = string("q_175")]; + tensor k_173 = transpose(perm = k_173_perm_0, x = k_171)[name = string("transpose_26")]; + tensor var_8750 = mul(x = k_173, y = cos_3)[name = string("op_8750")]; + tensor x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_87 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_173)[name = string("x1_87")]; + tensor x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_87 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_173)[name = string("x2_87")]; + fp32 const_219_promoted = const()[name = string("const_219_promoted"), val = fp32(-0x1p+0)]; + tensor var_8771 = mul(x = x2_87, y = const_219_promoted)[name = string("op_8771")]; + int32 var_8773 = const()[name = string("op_8773"), val = int32(-1)]; + bool var_8774_interleave_0 = const()[name = string("op_8774_interleave_0"), val = bool(false)]; + tensor var_8774 = concat(axis = var_8773, interleave = var_8774_interleave_0, values = (var_8771, x1_87))[name = string("op_8774")]; + tensor var_8775 = mul(x = var_8774, y = sin_3)[name = string("op_8775")]; + tensor k_175 = add(x = var_8750, y = var_8775)[name = string("k_175")]; + tensor var_8782 = const()[name = string("op_8782"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_43 = reshape(shape = var_8782, x = k_175)[name = string("nk_flat_43")]; + tensor var_8788 = const()[name = string("op_8788"), val = tensor([1, 1024, 1, 1])]; + tensor v_131 = transpose(perm = v_131_perm_0, x = v_129)[name = string("transpose_25")]; + tensor nv_flat_43 = reshape(shape = var_8788, x = v_131)[name = string("nv_flat_43")]; + tensor var_8797 = mul(x = var_8591, y = var_1194)[name = string("op_8797")]; + tensor var_8798 = mul(x = nk_flat_43, y = update_mask_1)[name = string("op_8798")]; + tensor key_cache_89 = add(x = var_8797, y = var_8798)[name = string("key_cache_89")]; + tensor var_8804 = mul(x = var_8611, y = var_1194)[name = string("op_8804")]; + tensor var_8805 = mul(x = nv_flat_43, y = update_mask_1)[name = string("op_8805")]; + tensor value_cache_89 = add(x = var_8804, y = var_8805)[name = string("value_cache_89")]; + tensor kc_127_axes_0 = const()[name = string("kc_127_axes_0"), val = tensor([2])]; + tensor kc_127 = squeeze(axes = kc_127_axes_0, x = key_cache_89)[name = string("kc_127")]; + tensor var_8814 = const()[name = string("op_8814"), val = tensor([1, 8, 128, 256])]; + tensor kc_129 = reshape(shape = var_8814, x = kc_127)[name = string("kc_129")]; + tensor vc_127_axes_0 = const()[name = string("vc_127_axes_0"), val = tensor([2])]; + tensor vc_127 = squeeze(axes = vc_127_axes_0, x = value_cache_89)[name = string("vc_127")]; + tensor var_8822 = const()[name = string("op_8822"), val = tensor([1, 8, 128, 256])]; + tensor vc_129 = reshape(shape = var_8822, x = vc_127)[name = string("vc_129")]; + tensor var_8825_axes_0 = const()[name = string("op_8825_axes_0"), val = tensor([2])]; + tensor var_8825 = expand_dims(axes = var_8825_axes_0, x = kc_129)[name = string("op_8825")]; + tensor var_8833_reps_0 = const()[name = string("op_8833_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_8833 = tile(reps = var_8833_reps_0, x = var_8825)[name = string("op_8833")]; + tensor var_8838 = const()[name = string("op_8838"), val = tensor([1, 16, 128, 256])]; + tensor kc_131 = reshape(shape = var_8838, x = var_8833)[name = string("kc_131")]; + tensor var_8841_axes_0 = const()[name = string("op_8841_axes_0"), val = tensor([2])]; + tensor var_8841 = expand_dims(axes = var_8841_axes_0, x = vc_129)[name = string("op_8841")]; + tensor var_8849_reps_0 = const()[name = string("op_8849_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_8849 = tile(reps = var_8849_reps_0, x = var_8841)[name = string("op_8849")]; + tensor var_8854 = const()[name = string("op_8854"), val = tensor([1, 16, 128, 256])]; + tensor vc_131 = reshape(shape = var_8854, x = var_8849)[name = string("vc_131")]; + bool var_8856_transpose_x_0 = const()[name = string("op_8856_transpose_x_0"), val = bool(false)]; + bool var_8856_transpose_y_0 = const()[name = string("op_8856_transpose_y_0"), val = bool(false)]; + tensor var_8856 = matmul(transpose_x = var_8856_transpose_x_0, transpose_y = var_8856_transpose_y_0, x = q_175, y = kc_131)[name = string("op_8856")]; + fp32 _inversed_attn_weights_169_y_0 = const()[name = string("_inversed_attn_weights_169_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_169 = mul(x = var_8856, y = _inversed_attn_weights_169_y_0)[name = string("_inversed_attn_weights_169")]; + tensor attn_weights_171 = add(x = _inversed_attn_weights_169, y = mask_1)[name = string("attn_weights_171")]; + int32 var_8870 = const()[name = string("op_8870"), val = int32(-1)]; + tensor attn_weights_175 = softmax(axis = var_8870, x = attn_weights_171)[name = string("attn_weights_175")]; + bool attn_output_85_transpose_x_1 = const()[name = string("attn_output_85_transpose_x_1"), val = bool(false)]; + bool attn_output_85_transpose_y_1 = const()[name = string("attn_output_85_transpose_y_1"), val = bool(true)]; + tensor attn_output_85 = matmul(transpose_x = attn_output_85_transpose_x_1, transpose_y = attn_output_85_transpose_y_1, x = attn_weights_175, y = vc_131)[name = string("attn_output_85")]; + tensor var_8879_perm_0 = const()[name = string("op_8879_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_8883 = const()[name = string("op_8883"), val = tensor([1, 1, -1])]; + tensor var_8879 = transpose(perm = var_8879_perm_0, x = attn_output_85)[name = string("transpose_24")]; + tensor input_213 = reshape(shape = var_8883, x = var_8879)[name = string("input_213")]; + tensor attn_output_87 = linear(bias = linear_1_bias_0, weight = layers_21_self_attn_o_proj_weight, x = input_213)[name = string("linear_150")]; + tensor var_8889_axes_0 = const()[name = string("op_8889_axes_0"), val = tensor([0])]; + tensor var_8889 = squeeze(axes = var_8889_axes_0, x = attn_output_87)[name = string("op_8889")]; + tensor var_8891_axes_0 = const()[name = string("op_8891_axes_0"), val = tensor([0])]; + tensor var_8891 = squeeze(axes = var_8891_axes_0, x = var_8889)[name = string("op_8891")]; + tensor var_8893_axes_0 = const()[name = string("op_8893_axes_0"), val = tensor([-1])]; + tensor var_8893 = expand_dims(axes = var_8893_axes_0, x = var_8891)[name = string("op_8893")]; + tensor attn_4d_43_axes_0 = const()[name = string("attn_4d_43_axes_0"), val = tensor([-1])]; + tensor attn_4d_43 = expand_dims(axes = attn_4d_43_axes_0, x = var_8893)[name = string("attn_4d_43")]; + tensor hidden_85 = add(x = hidden_83, y = attn_4d_43)[name = string("hidden_85")]; + tensor var_8899_axes_0 = const()[name = string("op_8899_axes_0"), val = tensor([-1])]; + tensor var_8899 = squeeze(axes = var_8899_axes_0, x = hidden_85)[name = string("op_8899")]; + tensor var_8901_axes_0 = const()[name = string("op_8901_axes_0"), val = tensor([-1])]; + tensor var_8901 = squeeze(axes = var_8901_axes_0, x = var_8899)[name = string("op_8901")]; + tensor hidden_states_523_axes_0 = const()[name = string("hidden_states_523_axes_0"), val = tensor([0])]; + tensor hidden_states_523 = expand_dims(axes = hidden_states_523_axes_0, x = var_8901)[name = string("hidden_states_523")]; + fp32 var_8907_promoted = const()[name = string("op_8907_promoted"), val = fp32(0x1p+1)]; + tensor var_8913 = pow(x = hidden_states_523, y = var_8907_promoted)[name = string("op_8913")]; + tensor variance_175_axes_0 = const()[name = string("variance_175_axes_0"), val = tensor([-1])]; + bool variance_175_keep_dims_0 = const()[name = string("variance_175_keep_dims_0"), val = bool(true)]; + tensor variance_175 = reduce_mean(axes = variance_175_axes_0, keep_dims = variance_175_keep_dims_0, x = var_8913)[name = string("variance_175")]; + fp32 var_8916 = const()[name = string("op_8916"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_8917 = add(x = variance_175, y = var_8916)[name = string("op_8917")]; + fp32 var_8918_epsilon_0 = const()[name = string("op_8918_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_8918 = rsqrt(epsilon = var_8918_epsilon_0, x = var_8917)[name = string("op_8918")]; + tensor hidden_states_527 = mul(x = hidden_states_523, y = var_8918)[name = string("hidden_states_527")]; + tensor const_220 = const()[name = string("const_220"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774432576)))]; + tensor input_215 = mul(x = const_220, y = hidden_states_527)[name = string("input_215")]; + tensor input_217 = linear(bias = linear_4_bias_0, weight = layers_21_mlp_gate_proj_weight, x = input_215)[name = string("linear_151")]; + tensor var_8928 = silu(x = input_217)[name = string("op_8928")]; + tensor var_8930 = linear(bias = linear_4_bias_0, weight = layers_21_mlp_up_proj_weight, x = input_215)[name = string("linear_152")]; + tensor input_219 = mul(x = var_8928, y = var_8930)[name = string("input_219")]; + tensor mlp_out_43 = linear(bias = linear_1_bias_0, weight = layers_21_mlp_down_proj_weight, x = input_219)[name = string("linear_153")]; + tensor var_8935_axes_0 = const()[name = string("op_8935_axes_0"), val = tensor([0])]; + tensor var_8935 = squeeze(axes = var_8935_axes_0, x = mlp_out_43)[name = string("op_8935")]; + tensor var_8937_axes_0 = const()[name = string("op_8937_axes_0"), val = tensor([0])]; + tensor var_8937 = squeeze(axes = var_8937_axes_0, x = var_8935)[name = string("op_8937")]; + tensor var_8939_axes_0 = const()[name = string("op_8939_axes_0"), val = tensor([-1])]; + tensor var_8939 = expand_dims(axes = var_8939_axes_0, x = var_8937)[name = string("op_8939")]; + tensor mlp_4d_43_axes_0 = const()[name = string("mlp_4d_43_axes_0"), val = tensor([-1])]; + tensor mlp_4d_43 = expand_dims(axes = mlp_4d_43_axes_0, x = var_8939)[name = string("mlp_4d_43")]; + tensor hidden_87 = add(x = hidden_85, y = mlp_4d_43)[name = string("hidden_87")]; + tensor var_8953_begin_0 = const()[name = string("op_8953_begin_0"), val = tensor([0, 22528, 0, 0])]; + tensor var_8953_end_0 = const()[name = string("op_8953_end_0"), val = tensor([1, 23552, 1, 256])]; + tensor var_8953_end_mask_0 = const()[name = string("op_8953_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_8953 = slice_by_index(begin = var_8953_begin_0, end = var_8953_end_0, end_mask = var_8953_end_mask_0, x = cast_3)[name = string("op_8953")]; + tensor var_8973_begin_0 = const()[name = string("op_8973_begin_0"), val = tensor([0, 22528, 0, 0])]; + tensor var_8973_end_0 = const()[name = string("op_8973_end_0"), val = tensor([1, 23552, 1, 256])]; + tensor var_8973_end_mask_0 = const()[name = string("op_8973_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_8973 = slice_by_index(begin = var_8973_begin_0, end = var_8973_end_0, end_mask = var_8973_end_mask_0, x = cast_4)[name = string("op_8973")]; + tensor var_8985_axes_0 = const()[name = string("op_8985_axes_0"), val = tensor([-1])]; + tensor var_8985 = squeeze(axes = var_8985_axes_0, x = hidden_87)[name = string("op_8985")]; + tensor var_8987_axes_0 = const()[name = string("op_8987_axes_0"), val = tensor([-1])]; + tensor var_8987 = squeeze(axes = var_8987_axes_0, x = var_8985)[name = string("op_8987")]; + tensor hidden_states_529_axes_0 = const()[name = string("hidden_states_529_axes_0"), val = tensor([0])]; + tensor hidden_states_529 = expand_dims(axes = hidden_states_529_axes_0, x = var_8987)[name = string("hidden_states_529")]; + fp32 var_8993_promoted = const()[name = string("op_8993_promoted"), val = fp32(0x1p+1)]; + tensor var_8999 = pow(x = hidden_states_529, y = var_8993_promoted)[name = string("op_8999")]; + tensor variance_177_axes_0 = const()[name = string("variance_177_axes_0"), val = tensor([-1])]; + bool variance_177_keep_dims_0 = const()[name = string("variance_177_keep_dims_0"), val = bool(true)]; + tensor variance_177 = reduce_mean(axes = variance_177_axes_0, keep_dims = variance_177_keep_dims_0, x = var_8999)[name = string("variance_177")]; + fp32 var_9002 = const()[name = string("op_9002"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9003 = add(x = variance_177, y = var_9002)[name = string("op_9003")]; + fp32 var_9004_epsilon_0 = const()[name = string("op_9004_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9004 = rsqrt(epsilon = var_9004_epsilon_0, x = var_9003)[name = string("op_9004")]; + tensor hidden_states_533 = mul(x = hidden_states_529, y = var_9004)[name = string("hidden_states_533")]; + tensor const_221 = const()[name = string("const_221"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774436736)))]; + tensor input_221 = mul(x = const_221, y = hidden_states_533)[name = string("input_221")]; + tensor q_177 = linear(bias = linear_0_bias_0, weight = layers_22_self_attn_q_proj_weight, x = input_221)[name = string("linear_154")]; + tensor k_177 = linear(bias = linear_1_bias_0, weight = layers_22_self_attn_k_proj_weight, x = input_221)[name = string("linear_155")]; + tensor v_133 = linear(bias = linear_1_bias_0, weight = layers_22_self_attn_v_proj_weight, x = input_221)[name = string("linear_156")]; + tensor var_9021 = const()[name = string("op_9021"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_535 = reshape(shape = var_9021, x = q_177)[name = string("hidden_states_535")]; + tensor var_9027 = const()[name = string("op_9027"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_541 = reshape(shape = var_9027, x = k_177)[name = string("hidden_states_541")]; + tensor var_9033 = const()[name = string("op_9033"), val = tensor([1, 1, 8, 128])]; + tensor v_135 = reshape(shape = var_9033, x = v_133)[name = string("v_135")]; + fp32 var_9038_promoted = const()[name = string("op_9038_promoted"), val = fp32(0x1p+1)]; + tensor var_9044 = pow(x = hidden_states_535, y = var_9038_promoted)[name = string("op_9044")]; + tensor variance_179_axes_0 = const()[name = string("variance_179_axes_0"), val = tensor([-1])]; + bool variance_179_keep_dims_0 = const()[name = string("variance_179_keep_dims_0"), val = bool(true)]; + tensor variance_179 = reduce_mean(axes = variance_179_axes_0, keep_dims = variance_179_keep_dims_0, x = var_9044)[name = string("variance_179")]; + fp32 var_9047 = const()[name = string("op_9047"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9048 = add(x = variance_179, y = var_9047)[name = string("op_9048")]; + fp32 var_9049_epsilon_0 = const()[name = string("op_9049_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9049 = rsqrt(epsilon = var_9049_epsilon_0, x = var_9048)[name = string("op_9049")]; + tensor hidden_states_539 = mul(x = hidden_states_535, y = var_9049)[name = string("hidden_states_539")]; + tensor const_222 = const()[name = string("const_222"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774440896)))]; + tensor q_179 = mul(x = const_222, y = hidden_states_539)[name = string("q_179")]; + fp32 var_9056_promoted = const()[name = string("op_9056_promoted"), val = fp32(0x1p+1)]; + tensor var_9062 = pow(x = hidden_states_541, y = var_9056_promoted)[name = string("op_9062")]; + tensor variance_181_axes_0 = const()[name = string("variance_181_axes_0"), val = tensor([-1])]; + bool variance_181_keep_dims_0 = const()[name = string("variance_181_keep_dims_0"), val = bool(true)]; + tensor variance_181 = reduce_mean(axes = variance_181_axes_0, keep_dims = variance_181_keep_dims_0, x = var_9062)[name = string("variance_181")]; + fp32 var_9065 = const()[name = string("op_9065"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9066 = add(x = variance_181, y = var_9065)[name = string("op_9066")]; + fp32 var_9067_epsilon_0 = const()[name = string("op_9067_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9067 = rsqrt(epsilon = var_9067_epsilon_0, x = var_9066)[name = string("op_9067")]; + tensor hidden_states_545 = mul(x = hidden_states_541, y = var_9067)[name = string("hidden_states_545")]; + tensor const_223 = const()[name = string("const_223"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774441472)))]; + tensor k_179 = mul(x = const_223, y = hidden_states_545)[name = string("k_179")]; + tensor q_181_perm_0 = const()[name = string("q_181_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_181_perm_0 = const()[name = string("k_181_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_137_perm_0 = const()[name = string("v_137_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_181 = transpose(perm = q_181_perm_0, x = q_179)[name = string("transpose_23")]; + tensor var_9084 = mul(x = q_181, y = cos_3)[name = string("op_9084")]; + tensor x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_89 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_181)[name = string("x1_89")]; + tensor x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_89 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_181)[name = string("x2_89")]; + fp32 const_226_promoted = const()[name = string("const_226_promoted"), val = fp32(-0x1p+0)]; + tensor var_9105 = mul(x = x2_89, y = const_226_promoted)[name = string("op_9105")]; + int32 var_9107 = const()[name = string("op_9107"), val = int32(-1)]; + bool var_9108_interleave_0 = const()[name = string("op_9108_interleave_0"), val = bool(false)]; + tensor var_9108 = concat(axis = var_9107, interleave = var_9108_interleave_0, values = (var_9105, x1_89))[name = string("op_9108")]; + tensor var_9109 = mul(x = var_9108, y = sin_3)[name = string("op_9109")]; + tensor q_183 = add(x = var_9084, y = var_9109)[name = string("q_183")]; + tensor k_181 = transpose(perm = k_181_perm_0, x = k_179)[name = string("transpose_22")]; + tensor var_9112 = mul(x = k_181, y = cos_3)[name = string("op_9112")]; + tensor x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_91 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_181)[name = string("x1_91")]; + tensor x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_91 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_181)[name = string("x2_91")]; + fp32 const_229_promoted = const()[name = string("const_229_promoted"), val = fp32(-0x1p+0)]; + tensor var_9133 = mul(x = x2_91, y = const_229_promoted)[name = string("op_9133")]; + int32 var_9135 = const()[name = string("op_9135"), val = int32(-1)]; + bool var_9136_interleave_0 = const()[name = string("op_9136_interleave_0"), val = bool(false)]; + tensor var_9136 = concat(axis = var_9135, interleave = var_9136_interleave_0, values = (var_9133, x1_91))[name = string("op_9136")]; + tensor var_9137 = mul(x = var_9136, y = sin_3)[name = string("op_9137")]; + tensor k_183 = add(x = var_9112, y = var_9137)[name = string("k_183")]; + tensor var_9144 = const()[name = string("op_9144"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_45 = reshape(shape = var_9144, x = k_183)[name = string("nk_flat_45")]; + tensor var_9150 = const()[name = string("op_9150"), val = tensor([1, 1024, 1, 1])]; + tensor v_137 = transpose(perm = v_137_perm_0, x = v_135)[name = string("transpose_21")]; + tensor nv_flat_45 = reshape(shape = var_9150, x = v_137)[name = string("nv_flat_45")]; + tensor var_9159 = mul(x = var_8953, y = var_1194)[name = string("op_9159")]; + tensor var_9160 = mul(x = nk_flat_45, y = update_mask_1)[name = string("op_9160")]; + tensor key_cache_93 = add(x = var_9159, y = var_9160)[name = string("key_cache_93")]; + tensor var_9166 = mul(x = var_8973, y = var_1194)[name = string("op_9166")]; + tensor var_9167 = mul(x = nv_flat_45, y = update_mask_1)[name = string("op_9167")]; + tensor value_cache_93 = add(x = var_9166, y = var_9167)[name = string("value_cache_93")]; + tensor kc_133_axes_0 = const()[name = string("kc_133_axes_0"), val = tensor([2])]; + tensor kc_133 = squeeze(axes = kc_133_axes_0, x = key_cache_93)[name = string("kc_133")]; + tensor var_9176 = const()[name = string("op_9176"), val = tensor([1, 8, 128, 256])]; + tensor kc_135 = reshape(shape = var_9176, x = kc_133)[name = string("kc_135")]; + tensor vc_133_axes_0 = const()[name = string("vc_133_axes_0"), val = tensor([2])]; + tensor vc_133 = squeeze(axes = vc_133_axes_0, x = value_cache_93)[name = string("vc_133")]; + tensor var_9184 = const()[name = string("op_9184"), val = tensor([1, 8, 128, 256])]; + tensor vc_135 = reshape(shape = var_9184, x = vc_133)[name = string("vc_135")]; + tensor var_9187_axes_0 = const()[name = string("op_9187_axes_0"), val = tensor([2])]; + tensor var_9187 = expand_dims(axes = var_9187_axes_0, x = kc_135)[name = string("op_9187")]; + tensor var_9195_reps_0 = const()[name = string("op_9195_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_9195 = tile(reps = var_9195_reps_0, x = var_9187)[name = string("op_9195")]; + tensor var_9200 = const()[name = string("op_9200"), val = tensor([1, 16, 128, 256])]; + tensor kc_137 = reshape(shape = var_9200, x = var_9195)[name = string("kc_137")]; + tensor var_9203_axes_0 = const()[name = string("op_9203_axes_0"), val = tensor([2])]; + tensor var_9203 = expand_dims(axes = var_9203_axes_0, x = vc_135)[name = string("op_9203")]; + tensor var_9211_reps_0 = const()[name = string("op_9211_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_9211 = tile(reps = var_9211_reps_0, x = var_9203)[name = string("op_9211")]; + tensor var_9216 = const()[name = string("op_9216"), val = tensor([1, 16, 128, 256])]; + tensor vc_137 = reshape(shape = var_9216, x = var_9211)[name = string("vc_137")]; + bool var_9218_transpose_x_0 = const()[name = string("op_9218_transpose_x_0"), val = bool(false)]; + bool var_9218_transpose_y_0 = const()[name = string("op_9218_transpose_y_0"), val = bool(false)]; + tensor var_9218 = matmul(transpose_x = var_9218_transpose_x_0, transpose_y = var_9218_transpose_y_0, x = q_183, y = kc_137)[name = string("op_9218")]; + fp32 _inversed_attn_weights_177_y_0 = const()[name = string("_inversed_attn_weights_177_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_177 = mul(x = var_9218, y = _inversed_attn_weights_177_y_0)[name = string("_inversed_attn_weights_177")]; + tensor attn_weights_179 = add(x = _inversed_attn_weights_177, y = mask_1)[name = string("attn_weights_179")]; + int32 var_9232 = const()[name = string("op_9232"), val = int32(-1)]; + tensor attn_weights_183 = softmax(axis = var_9232, x = attn_weights_179)[name = string("attn_weights_183")]; + bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(false)]; + bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(true)]; + tensor attn_output_89 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_183, y = vc_137)[name = string("attn_output_89")]; + tensor var_9241_perm_0 = const()[name = string("op_9241_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_9245 = const()[name = string("op_9245"), val = tensor([1, 1, -1])]; + tensor var_9241 = transpose(perm = var_9241_perm_0, x = attn_output_89)[name = string("transpose_20")]; + tensor input_223 = reshape(shape = var_9245, x = var_9241)[name = string("input_223")]; + tensor attn_output_91 = linear(bias = linear_1_bias_0, weight = layers_22_self_attn_o_proj_weight, x = input_223)[name = string("linear_157")]; + tensor var_9251_axes_0 = const()[name = string("op_9251_axes_0"), val = tensor([0])]; + tensor var_9251 = squeeze(axes = var_9251_axes_0, x = attn_output_91)[name = string("op_9251")]; + tensor var_9253_axes_0 = const()[name = string("op_9253_axes_0"), val = tensor([0])]; + tensor var_9253 = squeeze(axes = var_9253_axes_0, x = var_9251)[name = string("op_9253")]; + tensor var_9255_axes_0 = const()[name = string("op_9255_axes_0"), val = tensor([-1])]; + tensor var_9255 = expand_dims(axes = var_9255_axes_0, x = var_9253)[name = string("op_9255")]; + tensor attn_4d_45_axes_0 = const()[name = string("attn_4d_45_axes_0"), val = tensor([-1])]; + tensor attn_4d_45 = expand_dims(axes = attn_4d_45_axes_0, x = var_9255)[name = string("attn_4d_45")]; + tensor hidden_89 = add(x = hidden_87, y = attn_4d_45)[name = string("hidden_89")]; + tensor var_9261_axes_0 = const()[name = string("op_9261_axes_0"), val = tensor([-1])]; + tensor var_9261 = squeeze(axes = var_9261_axes_0, x = hidden_89)[name = string("op_9261")]; + tensor var_9263_axes_0 = const()[name = string("op_9263_axes_0"), val = tensor([-1])]; + tensor var_9263 = squeeze(axes = var_9263_axes_0, x = var_9261)[name = string("op_9263")]; + tensor hidden_states_547_axes_0 = const()[name = string("hidden_states_547_axes_0"), val = tensor([0])]; + tensor hidden_states_547 = expand_dims(axes = hidden_states_547_axes_0, x = var_9263)[name = string("hidden_states_547")]; + fp32 var_9269_promoted = const()[name = string("op_9269_promoted"), val = fp32(0x1p+1)]; + tensor var_9275 = pow(x = hidden_states_547, y = var_9269_promoted)[name = string("op_9275")]; + tensor variance_183_axes_0 = const()[name = string("variance_183_axes_0"), val = tensor([-1])]; + bool variance_183_keep_dims_0 = const()[name = string("variance_183_keep_dims_0"), val = bool(true)]; + tensor variance_183 = reduce_mean(axes = variance_183_axes_0, keep_dims = variance_183_keep_dims_0, x = var_9275)[name = string("variance_183")]; + fp32 var_9278 = const()[name = string("op_9278"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9279 = add(x = variance_183, y = var_9278)[name = string("op_9279")]; + fp32 var_9280_epsilon_0 = const()[name = string("op_9280_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9280 = rsqrt(epsilon = var_9280_epsilon_0, x = var_9279)[name = string("op_9280")]; + tensor hidden_states_551 = mul(x = hidden_states_547, y = var_9280)[name = string("hidden_states_551")]; + tensor const_230 = const()[name = string("const_230"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774442048)))]; + tensor input_225 = mul(x = const_230, y = hidden_states_551)[name = string("input_225")]; + tensor input_227 = linear(bias = linear_4_bias_0, weight = layers_22_mlp_gate_proj_weight, x = input_225)[name = string("linear_158")]; + tensor var_9290 = silu(x = input_227)[name = string("op_9290")]; + tensor var_9292 = linear(bias = linear_4_bias_0, weight = layers_22_mlp_up_proj_weight, x = input_225)[name = string("linear_159")]; + tensor input_229 = mul(x = var_9290, y = var_9292)[name = string("input_229")]; + tensor mlp_out_45 = linear(bias = linear_1_bias_0, weight = layers_22_mlp_down_proj_weight, x = input_229)[name = string("linear_160")]; + tensor var_9297_axes_0 = const()[name = string("op_9297_axes_0"), val = tensor([0])]; + tensor var_9297 = squeeze(axes = var_9297_axes_0, x = mlp_out_45)[name = string("op_9297")]; + tensor var_9299_axes_0 = const()[name = string("op_9299_axes_0"), val = tensor([0])]; + tensor var_9299 = squeeze(axes = var_9299_axes_0, x = var_9297)[name = string("op_9299")]; + tensor var_9301_axes_0 = const()[name = string("op_9301_axes_0"), val = tensor([-1])]; + tensor var_9301 = expand_dims(axes = var_9301_axes_0, x = var_9299)[name = string("op_9301")]; + tensor mlp_4d_45_axes_0 = const()[name = string("mlp_4d_45_axes_0"), val = tensor([-1])]; + tensor mlp_4d_45 = expand_dims(axes = mlp_4d_45_axes_0, x = var_9301)[name = string("mlp_4d_45")]; + tensor hidden_91 = add(x = hidden_89, y = mlp_4d_45)[name = string("hidden_91")]; + tensor var_9315_begin_0 = const()[name = string("op_9315_begin_0"), val = tensor([0, 23552, 0, 0])]; + tensor var_9315_end_0 = const()[name = string("op_9315_end_0"), val = tensor([1, 24576, 1, 256])]; + tensor var_9315_end_mask_0 = const()[name = string("op_9315_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_9315 = slice_by_index(begin = var_9315_begin_0, end = var_9315_end_0, end_mask = var_9315_end_mask_0, x = cast_3)[name = string("op_9315")]; + tensor var_9335_begin_0 = const()[name = string("op_9335_begin_0"), val = tensor([0, 23552, 0, 0])]; + tensor var_9335_end_0 = const()[name = string("op_9335_end_0"), val = tensor([1, 24576, 1, 256])]; + tensor var_9335_end_mask_0 = const()[name = string("op_9335_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_9335 = slice_by_index(begin = var_9335_begin_0, end = var_9335_end_0, end_mask = var_9335_end_mask_0, x = cast_4)[name = string("op_9335")]; + tensor var_9347_axes_0 = const()[name = string("op_9347_axes_0"), val = tensor([-1])]; + tensor var_9347 = squeeze(axes = var_9347_axes_0, x = hidden_91)[name = string("op_9347")]; + tensor var_9349_axes_0 = const()[name = string("op_9349_axes_0"), val = tensor([-1])]; + tensor var_9349 = squeeze(axes = var_9349_axes_0, x = var_9347)[name = string("op_9349")]; + tensor hidden_states_553_axes_0 = const()[name = string("hidden_states_553_axes_0"), val = tensor([0])]; + tensor hidden_states_553 = expand_dims(axes = hidden_states_553_axes_0, x = var_9349)[name = string("hidden_states_553")]; + fp32 var_9355_promoted = const()[name = string("op_9355_promoted"), val = fp32(0x1p+1)]; + tensor var_9361 = pow(x = hidden_states_553, y = var_9355_promoted)[name = string("op_9361")]; + tensor variance_185_axes_0 = const()[name = string("variance_185_axes_0"), val = tensor([-1])]; + bool variance_185_keep_dims_0 = const()[name = string("variance_185_keep_dims_0"), val = bool(true)]; + tensor variance_185 = reduce_mean(axes = variance_185_axes_0, keep_dims = variance_185_keep_dims_0, x = var_9361)[name = string("variance_185")]; + fp32 var_9364 = const()[name = string("op_9364"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9365 = add(x = variance_185, y = var_9364)[name = string("op_9365")]; + fp32 var_9366_epsilon_0 = const()[name = string("op_9366_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9366 = rsqrt(epsilon = var_9366_epsilon_0, x = var_9365)[name = string("op_9366")]; + tensor hidden_states_557 = mul(x = hidden_states_553, y = var_9366)[name = string("hidden_states_557")]; + tensor const_231 = const()[name = string("const_231"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774446208)))]; + tensor input_231 = mul(x = const_231, y = hidden_states_557)[name = string("input_231")]; + tensor q_185 = linear(bias = linear_0_bias_0, weight = layers_23_self_attn_q_proj_weight, x = input_231)[name = string("linear_161")]; + tensor k_185 = linear(bias = linear_1_bias_0, weight = layers_23_self_attn_k_proj_weight, x = input_231)[name = string("linear_162")]; + tensor v_139 = linear(bias = linear_1_bias_0, weight = layers_23_self_attn_v_proj_weight, x = input_231)[name = string("linear_163")]; + tensor var_9383 = const()[name = string("op_9383"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_559 = reshape(shape = var_9383, x = q_185)[name = string("hidden_states_559")]; + tensor var_9389 = const()[name = string("op_9389"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_565 = reshape(shape = var_9389, x = k_185)[name = string("hidden_states_565")]; + tensor var_9395 = const()[name = string("op_9395"), val = tensor([1, 1, 8, 128])]; + tensor v_141 = reshape(shape = var_9395, x = v_139)[name = string("v_141")]; + fp32 var_9400_promoted = const()[name = string("op_9400_promoted"), val = fp32(0x1p+1)]; + tensor var_9406 = pow(x = hidden_states_559, y = var_9400_promoted)[name = string("op_9406")]; + tensor variance_187_axes_0 = const()[name = string("variance_187_axes_0"), val = tensor([-1])]; + bool variance_187_keep_dims_0 = const()[name = string("variance_187_keep_dims_0"), val = bool(true)]; + tensor variance_187 = reduce_mean(axes = variance_187_axes_0, keep_dims = variance_187_keep_dims_0, x = var_9406)[name = string("variance_187")]; + fp32 var_9409 = const()[name = string("op_9409"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9410 = add(x = variance_187, y = var_9409)[name = string("op_9410")]; + fp32 var_9411_epsilon_0 = const()[name = string("op_9411_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9411 = rsqrt(epsilon = var_9411_epsilon_0, x = var_9410)[name = string("op_9411")]; + tensor hidden_states_563 = mul(x = hidden_states_559, y = var_9411)[name = string("hidden_states_563")]; + tensor const_232 = const()[name = string("const_232"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774450368)))]; + tensor q_187 = mul(x = const_232, y = hidden_states_563)[name = string("q_187")]; + fp32 var_9418_promoted = const()[name = string("op_9418_promoted"), val = fp32(0x1p+1)]; + tensor var_9424 = pow(x = hidden_states_565, y = var_9418_promoted)[name = string("op_9424")]; + tensor variance_189_axes_0 = const()[name = string("variance_189_axes_0"), val = tensor([-1])]; + bool variance_189_keep_dims_0 = const()[name = string("variance_189_keep_dims_0"), val = bool(true)]; + tensor variance_189 = reduce_mean(axes = variance_189_axes_0, keep_dims = variance_189_keep_dims_0, x = var_9424)[name = string("variance_189")]; + fp32 var_9427 = const()[name = string("op_9427"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9428 = add(x = variance_189, y = var_9427)[name = string("op_9428")]; + fp32 var_9429_epsilon_0 = const()[name = string("op_9429_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9429 = rsqrt(epsilon = var_9429_epsilon_0, x = var_9428)[name = string("op_9429")]; + tensor hidden_states_569 = mul(x = hidden_states_565, y = var_9429)[name = string("hidden_states_569")]; + tensor const_233 = const()[name = string("const_233"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774450944)))]; + tensor k_187 = mul(x = const_233, y = hidden_states_569)[name = string("k_187")]; + tensor q_189_perm_0 = const()[name = string("q_189_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_189_perm_0 = const()[name = string("k_189_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_143_perm_0 = const()[name = string("v_143_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_189 = transpose(perm = q_189_perm_0, x = q_187)[name = string("transpose_19")]; + tensor var_9446 = mul(x = q_189, y = cos_3)[name = string("op_9446")]; + tensor x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_93 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_189)[name = string("x1_93")]; + tensor x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_93 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_189)[name = string("x2_93")]; + fp32 const_236_promoted = const()[name = string("const_236_promoted"), val = fp32(-0x1p+0)]; + tensor var_9467 = mul(x = x2_93, y = const_236_promoted)[name = string("op_9467")]; + int32 var_9469 = const()[name = string("op_9469"), val = int32(-1)]; + bool var_9470_interleave_0 = const()[name = string("op_9470_interleave_0"), val = bool(false)]; + tensor var_9470 = concat(axis = var_9469, interleave = var_9470_interleave_0, values = (var_9467, x1_93))[name = string("op_9470")]; + tensor var_9471 = mul(x = var_9470, y = sin_3)[name = string("op_9471")]; + tensor q_191 = add(x = var_9446, y = var_9471)[name = string("q_191")]; + tensor k_189 = transpose(perm = k_189_perm_0, x = k_187)[name = string("transpose_18")]; + tensor var_9474 = mul(x = k_189, y = cos_3)[name = string("op_9474")]; + tensor x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_95 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_189)[name = string("x1_95")]; + tensor x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_95 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_189)[name = string("x2_95")]; + fp32 const_239_promoted = const()[name = string("const_239_promoted"), val = fp32(-0x1p+0)]; + tensor var_9495 = mul(x = x2_95, y = const_239_promoted)[name = string("op_9495")]; + int32 var_9497 = const()[name = string("op_9497"), val = int32(-1)]; + bool var_9498_interleave_0 = const()[name = string("op_9498_interleave_0"), val = bool(false)]; + tensor var_9498 = concat(axis = var_9497, interleave = var_9498_interleave_0, values = (var_9495, x1_95))[name = string("op_9498")]; + tensor var_9499 = mul(x = var_9498, y = sin_3)[name = string("op_9499")]; + tensor k_191 = add(x = var_9474, y = var_9499)[name = string("k_191")]; + tensor var_9506 = const()[name = string("op_9506"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_47 = reshape(shape = var_9506, x = k_191)[name = string("nk_flat_47")]; + tensor var_9512 = const()[name = string("op_9512"), val = tensor([1, 1024, 1, 1])]; + tensor v_143 = transpose(perm = v_143_perm_0, x = v_141)[name = string("transpose_17")]; + tensor nv_flat_47 = reshape(shape = var_9512, x = v_143)[name = string("nv_flat_47")]; + tensor var_9521 = mul(x = var_9315, y = var_1194)[name = string("op_9521")]; + tensor var_9522 = mul(x = nk_flat_47, y = update_mask_1)[name = string("op_9522")]; + tensor key_cache_97 = add(x = var_9521, y = var_9522)[name = string("key_cache_97")]; + tensor var_9528 = mul(x = var_9335, y = var_1194)[name = string("op_9528")]; + tensor var_9529 = mul(x = nv_flat_47, y = update_mask_1)[name = string("op_9529")]; + tensor value_cache_97 = add(x = var_9528, y = var_9529)[name = string("value_cache_97")]; + tensor kc_139_axes_0 = const()[name = string("kc_139_axes_0"), val = tensor([2])]; + tensor kc_139 = squeeze(axes = kc_139_axes_0, x = key_cache_97)[name = string("kc_139")]; + tensor var_9538 = const()[name = string("op_9538"), val = tensor([1, 8, 128, 256])]; + tensor kc_141 = reshape(shape = var_9538, x = kc_139)[name = string("kc_141")]; + tensor vc_139_axes_0 = const()[name = string("vc_139_axes_0"), val = tensor([2])]; + tensor vc_139 = squeeze(axes = vc_139_axes_0, x = value_cache_97)[name = string("vc_139")]; + tensor var_9546 = const()[name = string("op_9546"), val = tensor([1, 8, 128, 256])]; + tensor vc_141 = reshape(shape = var_9546, x = vc_139)[name = string("vc_141")]; + tensor var_9549_axes_0 = const()[name = string("op_9549_axes_0"), val = tensor([2])]; + tensor var_9549 = expand_dims(axes = var_9549_axes_0, x = kc_141)[name = string("op_9549")]; + tensor var_9557_reps_0 = const()[name = string("op_9557_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_9557 = tile(reps = var_9557_reps_0, x = var_9549)[name = string("op_9557")]; + tensor var_9562 = const()[name = string("op_9562"), val = tensor([1, 16, 128, 256])]; + tensor kc_143 = reshape(shape = var_9562, x = var_9557)[name = string("kc_143")]; + tensor var_9565_axes_0 = const()[name = string("op_9565_axes_0"), val = tensor([2])]; + tensor var_9565 = expand_dims(axes = var_9565_axes_0, x = vc_141)[name = string("op_9565")]; + tensor var_9573_reps_0 = const()[name = string("op_9573_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_9573 = tile(reps = var_9573_reps_0, x = var_9565)[name = string("op_9573")]; + tensor var_9578 = const()[name = string("op_9578"), val = tensor([1, 16, 128, 256])]; + tensor vc_143 = reshape(shape = var_9578, x = var_9573)[name = string("vc_143")]; + bool var_9580_transpose_x_0 = const()[name = string("op_9580_transpose_x_0"), val = bool(false)]; + bool var_9580_transpose_y_0 = const()[name = string("op_9580_transpose_y_0"), val = bool(false)]; + tensor var_9580 = matmul(transpose_x = var_9580_transpose_x_0, transpose_y = var_9580_transpose_y_0, x = q_191, y = kc_143)[name = string("op_9580")]; + fp32 _inversed_attn_weights_185_y_0 = const()[name = string("_inversed_attn_weights_185_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_185 = mul(x = var_9580, y = _inversed_attn_weights_185_y_0)[name = string("_inversed_attn_weights_185")]; + tensor attn_weights_187 = add(x = _inversed_attn_weights_185, y = mask_1)[name = string("attn_weights_187")]; + int32 var_9594 = const()[name = string("op_9594"), val = int32(-1)]; + tensor attn_weights_191 = softmax(axis = var_9594, x = attn_weights_187)[name = string("attn_weights_191")]; + bool attn_output_93_transpose_x_1 = const()[name = string("attn_output_93_transpose_x_1"), val = bool(false)]; + bool attn_output_93_transpose_y_1 = const()[name = string("attn_output_93_transpose_y_1"), val = bool(true)]; + tensor attn_output_93 = matmul(transpose_x = attn_output_93_transpose_x_1, transpose_y = attn_output_93_transpose_y_1, x = attn_weights_191, y = vc_143)[name = string("attn_output_93")]; + tensor var_9603_perm_0 = const()[name = string("op_9603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_9607 = const()[name = string("op_9607"), val = tensor([1, 1, -1])]; + tensor var_9603 = transpose(perm = var_9603_perm_0, x = attn_output_93)[name = string("transpose_16")]; + tensor input_233 = reshape(shape = var_9607, x = var_9603)[name = string("input_233")]; + tensor attn_output_95 = linear(bias = linear_1_bias_0, weight = layers_23_self_attn_o_proj_weight, x = input_233)[name = string("linear_164")]; + tensor var_9613_axes_0 = const()[name = string("op_9613_axes_0"), val = tensor([0])]; + tensor var_9613 = squeeze(axes = var_9613_axes_0, x = attn_output_95)[name = string("op_9613")]; + tensor var_9615_axes_0 = const()[name = string("op_9615_axes_0"), val = tensor([0])]; + tensor var_9615 = squeeze(axes = var_9615_axes_0, x = var_9613)[name = string("op_9615")]; + tensor var_9617_axes_0 = const()[name = string("op_9617_axes_0"), val = tensor([-1])]; + tensor var_9617 = expand_dims(axes = var_9617_axes_0, x = var_9615)[name = string("op_9617")]; + tensor attn_4d_47_axes_0 = const()[name = string("attn_4d_47_axes_0"), val = tensor([-1])]; + tensor attn_4d_47 = expand_dims(axes = attn_4d_47_axes_0, x = var_9617)[name = string("attn_4d_47")]; + tensor hidden_93 = add(x = hidden_91, y = attn_4d_47)[name = string("hidden_93")]; + tensor var_9623_axes_0 = const()[name = string("op_9623_axes_0"), val = tensor([-1])]; + tensor var_9623 = squeeze(axes = var_9623_axes_0, x = hidden_93)[name = string("op_9623")]; + tensor var_9625_axes_0 = const()[name = string("op_9625_axes_0"), val = tensor([-1])]; + tensor var_9625 = squeeze(axes = var_9625_axes_0, x = var_9623)[name = string("op_9625")]; + tensor hidden_states_571_axes_0 = const()[name = string("hidden_states_571_axes_0"), val = tensor([0])]; + tensor hidden_states_571 = expand_dims(axes = hidden_states_571_axes_0, x = var_9625)[name = string("hidden_states_571")]; + fp32 var_9631_promoted = const()[name = string("op_9631_promoted"), val = fp32(0x1p+1)]; + tensor var_9637 = pow(x = hidden_states_571, y = var_9631_promoted)[name = string("op_9637")]; + tensor variance_191_axes_0 = const()[name = string("variance_191_axes_0"), val = tensor([-1])]; + bool variance_191_keep_dims_0 = const()[name = string("variance_191_keep_dims_0"), val = bool(true)]; + tensor variance_191 = reduce_mean(axes = variance_191_axes_0, keep_dims = variance_191_keep_dims_0, x = var_9637)[name = string("variance_191")]; + fp32 var_9640 = const()[name = string("op_9640"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9641 = add(x = variance_191, y = var_9640)[name = string("op_9641")]; + fp32 var_9642_epsilon_0 = const()[name = string("op_9642_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9642 = rsqrt(epsilon = var_9642_epsilon_0, x = var_9641)[name = string("op_9642")]; + tensor hidden_states_575 = mul(x = hidden_states_571, y = var_9642)[name = string("hidden_states_575")]; + tensor const_240 = const()[name = string("const_240"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774451520)))]; + tensor input_235 = mul(x = const_240, y = hidden_states_575)[name = string("input_235")]; + tensor input_237 = linear(bias = linear_4_bias_0, weight = layers_23_mlp_gate_proj_weight, x = input_235)[name = string("linear_165")]; + tensor var_9652 = silu(x = input_237)[name = string("op_9652")]; + tensor var_9654 = linear(bias = linear_4_bias_0, weight = layers_23_mlp_up_proj_weight, x = input_235)[name = string("linear_166")]; + tensor input_239 = mul(x = var_9652, y = var_9654)[name = string("input_239")]; + tensor mlp_out_47 = linear(bias = linear_1_bias_0, weight = layers_23_mlp_down_proj_weight, x = input_239)[name = string("linear_167")]; + tensor var_9659_axes_0 = const()[name = string("op_9659_axes_0"), val = tensor([0])]; + tensor var_9659 = squeeze(axes = var_9659_axes_0, x = mlp_out_47)[name = string("op_9659")]; + tensor var_9661_axes_0 = const()[name = string("op_9661_axes_0"), val = tensor([0])]; + tensor var_9661 = squeeze(axes = var_9661_axes_0, x = var_9659)[name = string("op_9661")]; + tensor var_9663_axes_0 = const()[name = string("op_9663_axes_0"), val = tensor([-1])]; + tensor var_9663 = expand_dims(axes = var_9663_axes_0, x = var_9661)[name = string("op_9663")]; + tensor mlp_4d_47_axes_0 = const()[name = string("mlp_4d_47_axes_0"), val = tensor([-1])]; + tensor mlp_4d_47 = expand_dims(axes = mlp_4d_47_axes_0, x = var_9663)[name = string("mlp_4d_47")]; + tensor hidden_95 = add(x = hidden_93, y = mlp_4d_47)[name = string("hidden_95")]; + tensor var_9677_begin_0 = const()[name = string("op_9677_begin_0"), val = tensor([0, 24576, 0, 0])]; + tensor var_9677_end_0 = const()[name = string("op_9677_end_0"), val = tensor([1, 25600, 1, 256])]; + tensor var_9677_end_mask_0 = const()[name = string("op_9677_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_9677 = slice_by_index(begin = var_9677_begin_0, end = var_9677_end_0, end_mask = var_9677_end_mask_0, x = cast_3)[name = string("op_9677")]; + tensor var_9697_begin_0 = const()[name = string("op_9697_begin_0"), val = tensor([0, 24576, 0, 0])]; + tensor var_9697_end_0 = const()[name = string("op_9697_end_0"), val = tensor([1, 25600, 1, 256])]; + tensor var_9697_end_mask_0 = const()[name = string("op_9697_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_9697 = slice_by_index(begin = var_9697_begin_0, end = var_9697_end_0, end_mask = var_9697_end_mask_0, x = cast_4)[name = string("op_9697")]; + tensor var_9709_axes_0 = const()[name = string("op_9709_axes_0"), val = tensor([-1])]; + tensor var_9709 = squeeze(axes = var_9709_axes_0, x = hidden_95)[name = string("op_9709")]; + tensor var_9711_axes_0 = const()[name = string("op_9711_axes_0"), val = tensor([-1])]; + tensor var_9711 = squeeze(axes = var_9711_axes_0, x = var_9709)[name = string("op_9711")]; + tensor hidden_states_577_axes_0 = const()[name = string("hidden_states_577_axes_0"), val = tensor([0])]; + tensor hidden_states_577 = expand_dims(axes = hidden_states_577_axes_0, x = var_9711)[name = string("hidden_states_577")]; + fp32 var_9717_promoted = const()[name = string("op_9717_promoted"), val = fp32(0x1p+1)]; + tensor var_9723 = pow(x = hidden_states_577, y = var_9717_promoted)[name = string("op_9723")]; + tensor variance_193_axes_0 = const()[name = string("variance_193_axes_0"), val = tensor([-1])]; + bool variance_193_keep_dims_0 = const()[name = string("variance_193_keep_dims_0"), val = bool(true)]; + tensor variance_193 = reduce_mean(axes = variance_193_axes_0, keep_dims = variance_193_keep_dims_0, x = var_9723)[name = string("variance_193")]; + fp32 var_9726 = const()[name = string("op_9726"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9727 = add(x = variance_193, y = var_9726)[name = string("op_9727")]; + fp32 var_9728_epsilon_0 = const()[name = string("op_9728_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9728 = rsqrt(epsilon = var_9728_epsilon_0, x = var_9727)[name = string("op_9728")]; + tensor hidden_states_581 = mul(x = hidden_states_577, y = var_9728)[name = string("hidden_states_581")]; + tensor const_241 = const()[name = string("const_241"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774455680)))]; + tensor input_241 = mul(x = const_241, y = hidden_states_581)[name = string("input_241")]; + tensor q_193 = linear(bias = linear_0_bias_0, weight = layers_24_self_attn_q_proj_weight, x = input_241)[name = string("linear_168")]; + tensor k_193 = linear(bias = linear_1_bias_0, weight = layers_24_self_attn_k_proj_weight, x = input_241)[name = string("linear_169")]; + tensor v_145 = linear(bias = linear_1_bias_0, weight = layers_24_self_attn_v_proj_weight, x = input_241)[name = string("linear_170")]; + tensor var_9745 = const()[name = string("op_9745"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_583 = reshape(shape = var_9745, x = q_193)[name = string("hidden_states_583")]; + tensor var_9751 = const()[name = string("op_9751"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_589 = reshape(shape = var_9751, x = k_193)[name = string("hidden_states_589")]; + tensor var_9757 = const()[name = string("op_9757"), val = tensor([1, 1, 8, 128])]; + tensor v_147 = reshape(shape = var_9757, x = v_145)[name = string("v_147")]; + fp32 var_9762_promoted = const()[name = string("op_9762_promoted"), val = fp32(0x1p+1)]; + tensor var_9768 = pow(x = hidden_states_583, y = var_9762_promoted)[name = string("op_9768")]; + tensor variance_195_axes_0 = const()[name = string("variance_195_axes_0"), val = tensor([-1])]; + bool variance_195_keep_dims_0 = const()[name = string("variance_195_keep_dims_0"), val = bool(true)]; + tensor variance_195 = reduce_mean(axes = variance_195_axes_0, keep_dims = variance_195_keep_dims_0, x = var_9768)[name = string("variance_195")]; + fp32 var_9771 = const()[name = string("op_9771"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9772 = add(x = variance_195, y = var_9771)[name = string("op_9772")]; + fp32 var_9773_epsilon_0 = const()[name = string("op_9773_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9773 = rsqrt(epsilon = var_9773_epsilon_0, x = var_9772)[name = string("op_9773")]; + tensor hidden_states_587 = mul(x = hidden_states_583, y = var_9773)[name = string("hidden_states_587")]; + tensor const_242 = const()[name = string("const_242"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774459840)))]; + tensor q_195 = mul(x = const_242, y = hidden_states_587)[name = string("q_195")]; + fp32 var_9780_promoted = const()[name = string("op_9780_promoted"), val = fp32(0x1p+1)]; + tensor var_9786 = pow(x = hidden_states_589, y = var_9780_promoted)[name = string("op_9786")]; + tensor variance_197_axes_0 = const()[name = string("variance_197_axes_0"), val = tensor([-1])]; + bool variance_197_keep_dims_0 = const()[name = string("variance_197_keep_dims_0"), val = bool(true)]; + tensor variance_197 = reduce_mean(axes = variance_197_axes_0, keep_dims = variance_197_keep_dims_0, x = var_9786)[name = string("variance_197")]; + fp32 var_9789 = const()[name = string("op_9789"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_9790 = add(x = variance_197, y = var_9789)[name = string("op_9790")]; + fp32 var_9791_epsilon_0 = const()[name = string("op_9791_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_9791 = rsqrt(epsilon = var_9791_epsilon_0, x = var_9790)[name = string("op_9791")]; + tensor hidden_states_593 = mul(x = hidden_states_589, y = var_9791)[name = string("hidden_states_593")]; + tensor const_243 = const()[name = string("const_243"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774460416)))]; + tensor k_195 = mul(x = const_243, y = hidden_states_593)[name = string("k_195")]; + tensor q_197_perm_0 = const()[name = string("q_197_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_197_perm_0 = const()[name = string("k_197_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_149_perm_0 = const()[name = string("v_149_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_197 = transpose(perm = q_197_perm_0, x = q_195)[name = string("transpose_15")]; + tensor var_9808 = mul(x = q_197, y = cos_3)[name = string("op_9808")]; + tensor x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_97 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_197)[name = string("x1_97")]; + tensor x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_97 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_197)[name = string("x2_97")]; + fp32 const_246_promoted = const()[name = string("const_246_promoted"), val = fp32(-0x1p+0)]; + tensor var_9829 = mul(x = x2_97, y = const_246_promoted)[name = string("op_9829")]; + int32 var_9831 = const()[name = string("op_9831"), val = int32(-1)]; + bool var_9832_interleave_0 = const()[name = string("op_9832_interleave_0"), val = bool(false)]; + tensor var_9832 = concat(axis = var_9831, interleave = var_9832_interleave_0, values = (var_9829, x1_97))[name = string("op_9832")]; + tensor var_9833 = mul(x = var_9832, y = sin_3)[name = string("op_9833")]; + tensor q_199 = add(x = var_9808, y = var_9833)[name = string("q_199")]; + tensor k_197 = transpose(perm = k_197_perm_0, x = k_195)[name = string("transpose_14")]; + tensor var_9836 = mul(x = k_197, y = cos_3)[name = string("op_9836")]; + tensor x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_99 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_197)[name = string("x1_99")]; + tensor x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_99 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_197)[name = string("x2_99")]; + fp32 const_249_promoted = const()[name = string("const_249_promoted"), val = fp32(-0x1p+0)]; + tensor var_9857 = mul(x = x2_99, y = const_249_promoted)[name = string("op_9857")]; + int32 var_9859 = const()[name = string("op_9859"), val = int32(-1)]; + bool var_9860_interleave_0 = const()[name = string("op_9860_interleave_0"), val = bool(false)]; + tensor var_9860 = concat(axis = var_9859, interleave = var_9860_interleave_0, values = (var_9857, x1_99))[name = string("op_9860")]; + tensor var_9861 = mul(x = var_9860, y = sin_3)[name = string("op_9861")]; + tensor k_199 = add(x = var_9836, y = var_9861)[name = string("k_199")]; + tensor var_9868 = const()[name = string("op_9868"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_49 = reshape(shape = var_9868, x = k_199)[name = string("nk_flat_49")]; + tensor var_9874 = const()[name = string("op_9874"), val = tensor([1, 1024, 1, 1])]; + tensor v_149 = transpose(perm = v_149_perm_0, x = v_147)[name = string("transpose_13")]; + tensor nv_flat_49 = reshape(shape = var_9874, x = v_149)[name = string("nv_flat_49")]; + tensor var_9883 = mul(x = var_9677, y = var_1194)[name = string("op_9883")]; + tensor var_9884 = mul(x = nk_flat_49, y = update_mask_1)[name = string("op_9884")]; + tensor key_cache_101 = add(x = var_9883, y = var_9884)[name = string("key_cache_101")]; + tensor var_9890 = mul(x = var_9697, y = var_1194)[name = string("op_9890")]; + tensor var_9891 = mul(x = nv_flat_49, y = update_mask_1)[name = string("op_9891")]; + tensor value_cache_101 = add(x = var_9890, y = var_9891)[name = string("value_cache_101")]; + tensor kc_145_axes_0 = const()[name = string("kc_145_axes_0"), val = tensor([2])]; + tensor kc_145 = squeeze(axes = kc_145_axes_0, x = key_cache_101)[name = string("kc_145")]; + tensor var_9900 = const()[name = string("op_9900"), val = tensor([1, 8, 128, 256])]; + tensor kc_147 = reshape(shape = var_9900, x = kc_145)[name = string("kc_147")]; + tensor vc_145_axes_0 = const()[name = string("vc_145_axes_0"), val = tensor([2])]; + tensor vc_145 = squeeze(axes = vc_145_axes_0, x = value_cache_101)[name = string("vc_145")]; + tensor var_9908 = const()[name = string("op_9908"), val = tensor([1, 8, 128, 256])]; + tensor vc_147 = reshape(shape = var_9908, x = vc_145)[name = string("vc_147")]; + tensor var_9911_axes_0 = const()[name = string("op_9911_axes_0"), val = tensor([2])]; + tensor var_9911 = expand_dims(axes = var_9911_axes_0, x = kc_147)[name = string("op_9911")]; + tensor var_9919_reps_0 = const()[name = string("op_9919_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_9919 = tile(reps = var_9919_reps_0, x = var_9911)[name = string("op_9919")]; + tensor var_9924 = const()[name = string("op_9924"), val = tensor([1, 16, 128, 256])]; + tensor kc_149 = reshape(shape = var_9924, x = var_9919)[name = string("kc_149")]; + tensor var_9927_axes_0 = const()[name = string("op_9927_axes_0"), val = tensor([2])]; + tensor var_9927 = expand_dims(axes = var_9927_axes_0, x = vc_147)[name = string("op_9927")]; + tensor var_9935_reps_0 = const()[name = string("op_9935_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_9935 = tile(reps = var_9935_reps_0, x = var_9927)[name = string("op_9935")]; + tensor var_9940 = const()[name = string("op_9940"), val = tensor([1, 16, 128, 256])]; + tensor vc_149 = reshape(shape = var_9940, x = var_9935)[name = string("vc_149")]; + bool var_9942_transpose_x_0 = const()[name = string("op_9942_transpose_x_0"), val = bool(false)]; + bool var_9942_transpose_y_0 = const()[name = string("op_9942_transpose_y_0"), val = bool(false)]; + tensor var_9942 = matmul(transpose_x = var_9942_transpose_x_0, transpose_y = var_9942_transpose_y_0, x = q_199, y = kc_149)[name = string("op_9942")]; + fp32 _inversed_attn_weights_193_y_0 = const()[name = string("_inversed_attn_weights_193_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_193 = mul(x = var_9942, y = _inversed_attn_weights_193_y_0)[name = string("_inversed_attn_weights_193")]; + tensor attn_weights_195 = add(x = _inversed_attn_weights_193, y = mask_1)[name = string("attn_weights_195")]; + int32 var_9956 = const()[name = string("op_9956"), val = int32(-1)]; + tensor attn_weights_199 = softmax(axis = var_9956, x = attn_weights_195)[name = string("attn_weights_199")]; + bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(false)]; + bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(true)]; + tensor attn_output_97 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_199, y = vc_149)[name = string("attn_output_97")]; + tensor var_9965_perm_0 = const()[name = string("op_9965_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_9969 = const()[name = string("op_9969"), val = tensor([1, 1, -1])]; + tensor var_9965 = transpose(perm = var_9965_perm_0, x = attn_output_97)[name = string("transpose_12")]; + tensor input_243 = reshape(shape = var_9969, x = var_9965)[name = string("input_243")]; + tensor attn_output_99 = linear(bias = linear_1_bias_0, weight = layers_24_self_attn_o_proj_weight, x = input_243)[name = string("linear_171")]; + tensor var_9975_axes_0 = const()[name = string("op_9975_axes_0"), val = tensor([0])]; + tensor var_9975 = squeeze(axes = var_9975_axes_0, x = attn_output_99)[name = string("op_9975")]; + tensor var_9977_axes_0 = const()[name = string("op_9977_axes_0"), val = tensor([0])]; + tensor var_9977 = squeeze(axes = var_9977_axes_0, x = var_9975)[name = string("op_9977")]; + tensor var_9979_axes_0 = const()[name = string("op_9979_axes_0"), val = tensor([-1])]; + tensor var_9979 = expand_dims(axes = var_9979_axes_0, x = var_9977)[name = string("op_9979")]; + tensor attn_4d_49_axes_0 = const()[name = string("attn_4d_49_axes_0"), val = tensor([-1])]; + tensor attn_4d_49 = expand_dims(axes = attn_4d_49_axes_0, x = var_9979)[name = string("attn_4d_49")]; + tensor hidden_97 = add(x = hidden_95, y = attn_4d_49)[name = string("hidden_97")]; + tensor var_9985_axes_0 = const()[name = string("op_9985_axes_0"), val = tensor([-1])]; + tensor var_9985 = squeeze(axes = var_9985_axes_0, x = hidden_97)[name = string("op_9985")]; + tensor var_9987_axes_0 = const()[name = string("op_9987_axes_0"), val = tensor([-1])]; + tensor var_9987 = squeeze(axes = var_9987_axes_0, x = var_9985)[name = string("op_9987")]; + tensor hidden_states_595_axes_0 = const()[name = string("hidden_states_595_axes_0"), val = tensor([0])]; + tensor hidden_states_595 = expand_dims(axes = hidden_states_595_axes_0, x = var_9987)[name = string("hidden_states_595")]; + fp32 var_9993_promoted = const()[name = string("op_9993_promoted"), val = fp32(0x1p+1)]; + tensor var_9999 = pow(x = hidden_states_595, y = var_9993_promoted)[name = string("op_9999")]; + tensor variance_199_axes_0 = const()[name = string("variance_199_axes_0"), val = tensor([-1])]; + bool variance_199_keep_dims_0 = const()[name = string("variance_199_keep_dims_0"), val = bool(true)]; + tensor variance_199 = reduce_mean(axes = variance_199_axes_0, keep_dims = variance_199_keep_dims_0, x = var_9999)[name = string("variance_199")]; + fp32 var_10002 = const()[name = string("op_10002"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10003 = add(x = variance_199, y = var_10002)[name = string("op_10003")]; + fp32 var_10004_epsilon_0 = const()[name = string("op_10004_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10004 = rsqrt(epsilon = var_10004_epsilon_0, x = var_10003)[name = string("op_10004")]; + tensor hidden_states_599 = mul(x = hidden_states_595, y = var_10004)[name = string("hidden_states_599")]; + tensor const_250 = const()[name = string("const_250"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774460992)))]; + tensor input_245 = mul(x = const_250, y = hidden_states_599)[name = string("input_245")]; + tensor input_247 = linear(bias = linear_4_bias_0, weight = layers_24_mlp_gate_proj_weight, x = input_245)[name = string("linear_172")]; + tensor var_10014 = silu(x = input_247)[name = string("op_10014")]; + tensor var_10016 = linear(bias = linear_4_bias_0, weight = layers_24_mlp_up_proj_weight, x = input_245)[name = string("linear_173")]; + tensor input_249 = mul(x = var_10014, y = var_10016)[name = string("input_249")]; + tensor mlp_out_49 = linear(bias = linear_1_bias_0, weight = layers_24_mlp_down_proj_weight, x = input_249)[name = string("linear_174")]; + tensor var_10021_axes_0 = const()[name = string("op_10021_axes_0"), val = tensor([0])]; + tensor var_10021 = squeeze(axes = var_10021_axes_0, x = mlp_out_49)[name = string("op_10021")]; + tensor var_10023_axes_0 = const()[name = string("op_10023_axes_0"), val = tensor([0])]; + tensor var_10023 = squeeze(axes = var_10023_axes_0, x = var_10021)[name = string("op_10023")]; + tensor var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor([-1])]; + tensor var_10025 = expand_dims(axes = var_10025_axes_0, x = var_10023)[name = string("op_10025")]; + tensor mlp_4d_49_axes_0 = const()[name = string("mlp_4d_49_axes_0"), val = tensor([-1])]; + tensor mlp_4d_49 = expand_dims(axes = mlp_4d_49_axes_0, x = var_10025)[name = string("mlp_4d_49")]; + tensor hidden_99 = add(x = hidden_97, y = mlp_4d_49)[name = string("hidden_99")]; + tensor var_10039_begin_0 = const()[name = string("op_10039_begin_0"), val = tensor([0, 25600, 0, 0])]; + tensor var_10039_end_0 = const()[name = string("op_10039_end_0"), val = tensor([1, 26624, 1, 256])]; + tensor var_10039_end_mask_0 = const()[name = string("op_10039_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_10039 = slice_by_index(begin = var_10039_begin_0, end = var_10039_end_0, end_mask = var_10039_end_mask_0, x = cast_3)[name = string("op_10039")]; + tensor var_10059_begin_0 = const()[name = string("op_10059_begin_0"), val = tensor([0, 25600, 0, 0])]; + tensor var_10059_end_0 = const()[name = string("op_10059_end_0"), val = tensor([1, 26624, 1, 256])]; + tensor var_10059_end_mask_0 = const()[name = string("op_10059_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_10059 = slice_by_index(begin = var_10059_begin_0, end = var_10059_end_0, end_mask = var_10059_end_mask_0, x = cast_4)[name = string("op_10059")]; + tensor var_10071_axes_0 = const()[name = string("op_10071_axes_0"), val = tensor([-1])]; + tensor var_10071 = squeeze(axes = var_10071_axes_0, x = hidden_99)[name = string("op_10071")]; + tensor var_10073_axes_0 = const()[name = string("op_10073_axes_0"), val = tensor([-1])]; + tensor var_10073 = squeeze(axes = var_10073_axes_0, x = var_10071)[name = string("op_10073")]; + tensor hidden_states_601_axes_0 = const()[name = string("hidden_states_601_axes_0"), val = tensor([0])]; + tensor hidden_states_601 = expand_dims(axes = hidden_states_601_axes_0, x = var_10073)[name = string("hidden_states_601")]; + fp32 var_10079_promoted = const()[name = string("op_10079_promoted"), val = fp32(0x1p+1)]; + tensor var_10085 = pow(x = hidden_states_601, y = var_10079_promoted)[name = string("op_10085")]; + tensor variance_201_axes_0 = const()[name = string("variance_201_axes_0"), val = tensor([-1])]; + bool variance_201_keep_dims_0 = const()[name = string("variance_201_keep_dims_0"), val = bool(true)]; + tensor variance_201 = reduce_mean(axes = variance_201_axes_0, keep_dims = variance_201_keep_dims_0, x = var_10085)[name = string("variance_201")]; + fp32 var_10088 = const()[name = string("op_10088"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10089 = add(x = variance_201, y = var_10088)[name = string("op_10089")]; + fp32 var_10090_epsilon_0 = const()[name = string("op_10090_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10090 = rsqrt(epsilon = var_10090_epsilon_0, x = var_10089)[name = string("op_10090")]; + tensor hidden_states_605 = mul(x = hidden_states_601, y = var_10090)[name = string("hidden_states_605")]; + tensor const_251 = const()[name = string("const_251"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774465152)))]; + tensor input_251 = mul(x = const_251, y = hidden_states_605)[name = string("input_251")]; + tensor q_201 = linear(bias = linear_0_bias_0, weight = layers_25_self_attn_q_proj_weight, x = input_251)[name = string("linear_175")]; + tensor k_201 = linear(bias = linear_1_bias_0, weight = layers_25_self_attn_k_proj_weight, x = input_251)[name = string("linear_176")]; + tensor v_151 = linear(bias = linear_1_bias_0, weight = layers_25_self_attn_v_proj_weight, x = input_251)[name = string("linear_177")]; + tensor var_10107 = const()[name = string("op_10107"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_607 = reshape(shape = var_10107, x = q_201)[name = string("hidden_states_607")]; + tensor var_10113 = const()[name = string("op_10113"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_613 = reshape(shape = var_10113, x = k_201)[name = string("hidden_states_613")]; + tensor var_10119 = const()[name = string("op_10119"), val = tensor([1, 1, 8, 128])]; + tensor v_153 = reshape(shape = var_10119, x = v_151)[name = string("v_153")]; + fp32 var_10124_promoted = const()[name = string("op_10124_promoted"), val = fp32(0x1p+1)]; + tensor var_10130 = pow(x = hidden_states_607, y = var_10124_promoted)[name = string("op_10130")]; + tensor variance_203_axes_0 = const()[name = string("variance_203_axes_0"), val = tensor([-1])]; + bool variance_203_keep_dims_0 = const()[name = string("variance_203_keep_dims_0"), val = bool(true)]; + tensor variance_203 = reduce_mean(axes = variance_203_axes_0, keep_dims = variance_203_keep_dims_0, x = var_10130)[name = string("variance_203")]; + fp32 var_10133 = const()[name = string("op_10133"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10134 = add(x = variance_203, y = var_10133)[name = string("op_10134")]; + fp32 var_10135_epsilon_0 = const()[name = string("op_10135_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10135 = rsqrt(epsilon = var_10135_epsilon_0, x = var_10134)[name = string("op_10135")]; + tensor hidden_states_611 = mul(x = hidden_states_607, y = var_10135)[name = string("hidden_states_611")]; + tensor const_252 = const()[name = string("const_252"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774469312)))]; + tensor q_203 = mul(x = const_252, y = hidden_states_611)[name = string("q_203")]; + fp32 var_10142_promoted = const()[name = string("op_10142_promoted"), val = fp32(0x1p+1)]; + tensor var_10148 = pow(x = hidden_states_613, y = var_10142_promoted)[name = string("op_10148")]; + tensor variance_205_axes_0 = const()[name = string("variance_205_axes_0"), val = tensor([-1])]; + bool variance_205_keep_dims_0 = const()[name = string("variance_205_keep_dims_0"), val = bool(true)]; + tensor variance_205 = reduce_mean(axes = variance_205_axes_0, keep_dims = variance_205_keep_dims_0, x = var_10148)[name = string("variance_205")]; + fp32 var_10151 = const()[name = string("op_10151"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10152 = add(x = variance_205, y = var_10151)[name = string("op_10152")]; + fp32 var_10153_epsilon_0 = const()[name = string("op_10153_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10153 = rsqrt(epsilon = var_10153_epsilon_0, x = var_10152)[name = string("op_10153")]; + tensor hidden_states_617 = mul(x = hidden_states_613, y = var_10153)[name = string("hidden_states_617")]; + tensor const_253 = const()[name = string("const_253"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774469888)))]; + tensor k_203 = mul(x = const_253, y = hidden_states_617)[name = string("k_203")]; + tensor q_205_perm_0 = const()[name = string("q_205_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_205_perm_0 = const()[name = string("k_205_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_155_perm_0 = const()[name = string("v_155_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_205 = transpose(perm = q_205_perm_0, x = q_203)[name = string("transpose_11")]; + tensor var_10170 = mul(x = q_205, y = cos_3)[name = string("op_10170")]; + tensor x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_101 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_205)[name = string("x1_101")]; + tensor x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_101 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_205)[name = string("x2_101")]; + fp32 const_256_promoted = const()[name = string("const_256_promoted"), val = fp32(-0x1p+0)]; + tensor var_10191 = mul(x = x2_101, y = const_256_promoted)[name = string("op_10191")]; + int32 var_10193 = const()[name = string("op_10193"), val = int32(-1)]; + bool var_10194_interleave_0 = const()[name = string("op_10194_interleave_0"), val = bool(false)]; + tensor var_10194 = concat(axis = var_10193, interleave = var_10194_interleave_0, values = (var_10191, x1_101))[name = string("op_10194")]; + tensor var_10195 = mul(x = var_10194, y = sin_3)[name = string("op_10195")]; + tensor q_207 = add(x = var_10170, y = var_10195)[name = string("q_207")]; + tensor k_205 = transpose(perm = k_205_perm_0, x = k_203)[name = string("transpose_10")]; + tensor var_10198 = mul(x = k_205, y = cos_3)[name = string("op_10198")]; + tensor x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_103 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_205)[name = string("x1_103")]; + tensor x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_103 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_205)[name = string("x2_103")]; + fp32 const_259_promoted = const()[name = string("const_259_promoted"), val = fp32(-0x1p+0)]; + tensor var_10219 = mul(x = x2_103, y = const_259_promoted)[name = string("op_10219")]; + int32 var_10221 = const()[name = string("op_10221"), val = int32(-1)]; + bool var_10222_interleave_0 = const()[name = string("op_10222_interleave_0"), val = bool(false)]; + tensor var_10222 = concat(axis = var_10221, interleave = var_10222_interleave_0, values = (var_10219, x1_103))[name = string("op_10222")]; + tensor var_10223 = mul(x = var_10222, y = sin_3)[name = string("op_10223")]; + tensor k_207 = add(x = var_10198, y = var_10223)[name = string("k_207")]; + tensor var_10230 = const()[name = string("op_10230"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_51 = reshape(shape = var_10230, x = k_207)[name = string("nk_flat_51")]; + tensor var_10236 = const()[name = string("op_10236"), val = tensor([1, 1024, 1, 1])]; + tensor v_155 = transpose(perm = v_155_perm_0, x = v_153)[name = string("transpose_9")]; + tensor nv_flat_51 = reshape(shape = var_10236, x = v_155)[name = string("nv_flat_51")]; + tensor var_10245 = mul(x = var_10039, y = var_1194)[name = string("op_10245")]; + tensor var_10246 = mul(x = nk_flat_51, y = update_mask_1)[name = string("op_10246")]; + tensor key_cache_105 = add(x = var_10245, y = var_10246)[name = string("key_cache_105")]; + tensor var_10252 = mul(x = var_10059, y = var_1194)[name = string("op_10252")]; + tensor var_10253 = mul(x = nv_flat_51, y = update_mask_1)[name = string("op_10253")]; + tensor value_cache_105 = add(x = var_10252, y = var_10253)[name = string("value_cache_105")]; + tensor kc_151_axes_0 = const()[name = string("kc_151_axes_0"), val = tensor([2])]; + tensor kc_151 = squeeze(axes = kc_151_axes_0, x = key_cache_105)[name = string("kc_151")]; + tensor var_10262 = const()[name = string("op_10262"), val = tensor([1, 8, 128, 256])]; + tensor kc_153 = reshape(shape = var_10262, x = kc_151)[name = string("kc_153")]; + tensor vc_151_axes_0 = const()[name = string("vc_151_axes_0"), val = tensor([2])]; + tensor vc_151 = squeeze(axes = vc_151_axes_0, x = value_cache_105)[name = string("vc_151")]; + tensor var_10270 = const()[name = string("op_10270"), val = tensor([1, 8, 128, 256])]; + tensor vc_153 = reshape(shape = var_10270, x = vc_151)[name = string("vc_153")]; + tensor var_10273_axes_0 = const()[name = string("op_10273_axes_0"), val = tensor([2])]; + tensor var_10273 = expand_dims(axes = var_10273_axes_0, x = kc_153)[name = string("op_10273")]; + tensor var_10281_reps_0 = const()[name = string("op_10281_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_10281 = tile(reps = var_10281_reps_0, x = var_10273)[name = string("op_10281")]; + tensor var_10286 = const()[name = string("op_10286"), val = tensor([1, 16, 128, 256])]; + tensor kc_155 = reshape(shape = var_10286, x = var_10281)[name = string("kc_155")]; + tensor var_10289_axes_0 = const()[name = string("op_10289_axes_0"), val = tensor([2])]; + tensor var_10289 = expand_dims(axes = var_10289_axes_0, x = vc_153)[name = string("op_10289")]; + tensor var_10297_reps_0 = const()[name = string("op_10297_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_10297 = tile(reps = var_10297_reps_0, x = var_10289)[name = string("op_10297")]; + tensor var_10302 = const()[name = string("op_10302"), val = tensor([1, 16, 128, 256])]; + tensor vc_155 = reshape(shape = var_10302, x = var_10297)[name = string("vc_155")]; + bool var_10304_transpose_x_0 = const()[name = string("op_10304_transpose_x_0"), val = bool(false)]; + bool var_10304_transpose_y_0 = const()[name = string("op_10304_transpose_y_0"), val = bool(false)]; + tensor var_10304 = matmul(transpose_x = var_10304_transpose_x_0, transpose_y = var_10304_transpose_y_0, x = q_207, y = kc_155)[name = string("op_10304")]; + fp32 _inversed_attn_weights_201_y_0 = const()[name = string("_inversed_attn_weights_201_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_201 = mul(x = var_10304, y = _inversed_attn_weights_201_y_0)[name = string("_inversed_attn_weights_201")]; + tensor attn_weights_203 = add(x = _inversed_attn_weights_201, y = mask_1)[name = string("attn_weights_203")]; + int32 var_10318 = const()[name = string("op_10318"), val = int32(-1)]; + tensor attn_weights_207 = softmax(axis = var_10318, x = attn_weights_203)[name = string("attn_weights_207")]; + bool attn_output_101_transpose_x_1 = const()[name = string("attn_output_101_transpose_x_1"), val = bool(false)]; + bool attn_output_101_transpose_y_1 = const()[name = string("attn_output_101_transpose_y_1"), val = bool(true)]; + tensor attn_output_101 = matmul(transpose_x = attn_output_101_transpose_x_1, transpose_y = attn_output_101_transpose_y_1, x = attn_weights_207, y = vc_155)[name = string("attn_output_101")]; + tensor var_10327_perm_0 = const()[name = string("op_10327_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_10331 = const()[name = string("op_10331"), val = tensor([1, 1, -1])]; + tensor var_10327 = transpose(perm = var_10327_perm_0, x = attn_output_101)[name = string("transpose_8")]; + tensor input_253 = reshape(shape = var_10331, x = var_10327)[name = string("input_253")]; + tensor attn_output_103 = linear(bias = linear_1_bias_0, weight = layers_25_self_attn_o_proj_weight, x = input_253)[name = string("linear_178")]; + tensor var_10337_axes_0 = const()[name = string("op_10337_axes_0"), val = tensor([0])]; + tensor var_10337 = squeeze(axes = var_10337_axes_0, x = attn_output_103)[name = string("op_10337")]; + tensor var_10339_axes_0 = const()[name = string("op_10339_axes_0"), val = tensor([0])]; + tensor var_10339 = squeeze(axes = var_10339_axes_0, x = var_10337)[name = string("op_10339")]; + tensor var_10341_axes_0 = const()[name = string("op_10341_axes_0"), val = tensor([-1])]; + tensor var_10341 = expand_dims(axes = var_10341_axes_0, x = var_10339)[name = string("op_10341")]; + tensor attn_4d_51_axes_0 = const()[name = string("attn_4d_51_axes_0"), val = tensor([-1])]; + tensor attn_4d_51 = expand_dims(axes = attn_4d_51_axes_0, x = var_10341)[name = string("attn_4d_51")]; + tensor hidden_101 = add(x = hidden_99, y = attn_4d_51)[name = string("hidden_101")]; + tensor var_10347_axes_0 = const()[name = string("op_10347_axes_0"), val = tensor([-1])]; + tensor var_10347 = squeeze(axes = var_10347_axes_0, x = hidden_101)[name = string("op_10347")]; + tensor var_10349_axes_0 = const()[name = string("op_10349_axes_0"), val = tensor([-1])]; + tensor var_10349 = squeeze(axes = var_10349_axes_0, x = var_10347)[name = string("op_10349")]; + tensor hidden_states_619_axes_0 = const()[name = string("hidden_states_619_axes_0"), val = tensor([0])]; + tensor hidden_states_619 = expand_dims(axes = hidden_states_619_axes_0, x = var_10349)[name = string("hidden_states_619")]; + fp32 var_10355_promoted = const()[name = string("op_10355_promoted"), val = fp32(0x1p+1)]; + tensor var_10361 = pow(x = hidden_states_619, y = var_10355_promoted)[name = string("op_10361")]; + tensor variance_207_axes_0 = const()[name = string("variance_207_axes_0"), val = tensor([-1])]; + bool variance_207_keep_dims_0 = const()[name = string("variance_207_keep_dims_0"), val = bool(true)]; + tensor variance_207 = reduce_mean(axes = variance_207_axes_0, keep_dims = variance_207_keep_dims_0, x = var_10361)[name = string("variance_207")]; + fp32 var_10364 = const()[name = string("op_10364"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10365 = add(x = variance_207, y = var_10364)[name = string("op_10365")]; + fp32 var_10366_epsilon_0 = const()[name = string("op_10366_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10366 = rsqrt(epsilon = var_10366_epsilon_0, x = var_10365)[name = string("op_10366")]; + tensor hidden_states_623 = mul(x = hidden_states_619, y = var_10366)[name = string("hidden_states_623")]; + tensor const_260 = const()[name = string("const_260"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774470464)))]; + tensor input_255 = mul(x = const_260, y = hidden_states_623)[name = string("input_255")]; + tensor input_257 = linear(bias = linear_4_bias_0, weight = layers_25_mlp_gate_proj_weight, x = input_255)[name = string("linear_179")]; + tensor var_10376 = silu(x = input_257)[name = string("op_10376")]; + tensor var_10378 = linear(bias = linear_4_bias_0, weight = layers_25_mlp_up_proj_weight, x = input_255)[name = string("linear_180")]; + tensor input_259 = mul(x = var_10376, y = var_10378)[name = string("input_259")]; + tensor mlp_out_51 = linear(bias = linear_1_bias_0, weight = layers_25_mlp_down_proj_weight, x = input_259)[name = string("linear_181")]; + tensor var_10383_axes_0 = const()[name = string("op_10383_axes_0"), val = tensor([0])]; + tensor var_10383 = squeeze(axes = var_10383_axes_0, x = mlp_out_51)[name = string("op_10383")]; + tensor var_10385_axes_0 = const()[name = string("op_10385_axes_0"), val = tensor([0])]; + tensor var_10385 = squeeze(axes = var_10385_axes_0, x = var_10383)[name = string("op_10385")]; + tensor var_10387_axes_0 = const()[name = string("op_10387_axes_0"), val = tensor([-1])]; + tensor var_10387 = expand_dims(axes = var_10387_axes_0, x = var_10385)[name = string("op_10387")]; + tensor mlp_4d_51_axes_0 = const()[name = string("mlp_4d_51_axes_0"), val = tensor([-1])]; + tensor mlp_4d_51 = expand_dims(axes = mlp_4d_51_axes_0, x = var_10387)[name = string("mlp_4d_51")]; + tensor hidden_103 = add(x = hidden_101, y = mlp_4d_51)[name = string("hidden_103")]; + tensor var_10401_begin_0 = const()[name = string("op_10401_begin_0"), val = tensor([0, 26624, 0, 0])]; + tensor var_10401_end_0 = const()[name = string("op_10401_end_0"), val = tensor([1, 27648, 1, 256])]; + tensor var_10401_end_mask_0 = const()[name = string("op_10401_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_10401 = slice_by_index(begin = var_10401_begin_0, end = var_10401_end_0, end_mask = var_10401_end_mask_0, x = cast_3)[name = string("op_10401")]; + tensor var_10421_begin_0 = const()[name = string("op_10421_begin_0"), val = tensor([0, 26624, 0, 0])]; + tensor var_10421_end_0 = const()[name = string("op_10421_end_0"), val = tensor([1, 27648, 1, 256])]; + tensor var_10421_end_mask_0 = const()[name = string("op_10421_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_10421 = slice_by_index(begin = var_10421_begin_0, end = var_10421_end_0, end_mask = var_10421_end_mask_0, x = cast_4)[name = string("op_10421")]; + tensor var_10433_axes_0 = const()[name = string("op_10433_axes_0"), val = tensor([-1])]; + tensor var_10433 = squeeze(axes = var_10433_axes_0, x = hidden_103)[name = string("op_10433")]; + tensor var_10435_axes_0 = const()[name = string("op_10435_axes_0"), val = tensor([-1])]; + tensor var_10435 = squeeze(axes = var_10435_axes_0, x = var_10433)[name = string("op_10435")]; + tensor hidden_states_625_axes_0 = const()[name = string("hidden_states_625_axes_0"), val = tensor([0])]; + tensor hidden_states_625 = expand_dims(axes = hidden_states_625_axes_0, x = var_10435)[name = string("hidden_states_625")]; + fp32 var_10441_promoted = const()[name = string("op_10441_promoted"), val = fp32(0x1p+1)]; + tensor var_10447 = pow(x = hidden_states_625, y = var_10441_promoted)[name = string("op_10447")]; + tensor variance_209_axes_0 = const()[name = string("variance_209_axes_0"), val = tensor([-1])]; + bool variance_209_keep_dims_0 = const()[name = string("variance_209_keep_dims_0"), val = bool(true)]; + tensor variance_209 = reduce_mean(axes = variance_209_axes_0, keep_dims = variance_209_keep_dims_0, x = var_10447)[name = string("variance_209")]; + fp32 var_10450 = const()[name = string("op_10450"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10451 = add(x = variance_209, y = var_10450)[name = string("op_10451")]; + fp32 var_10452_epsilon_0 = const()[name = string("op_10452_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10452 = rsqrt(epsilon = var_10452_epsilon_0, x = var_10451)[name = string("op_10452")]; + tensor hidden_states_629 = mul(x = hidden_states_625, y = var_10452)[name = string("hidden_states_629")]; + tensor const_261 = const()[name = string("const_261"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774474624)))]; + tensor input_261 = mul(x = const_261, y = hidden_states_629)[name = string("input_261")]; + tensor q_209 = linear(bias = linear_0_bias_0, weight = layers_26_self_attn_q_proj_weight, x = input_261)[name = string("linear_182")]; + tensor k_209 = linear(bias = linear_1_bias_0, weight = layers_26_self_attn_k_proj_weight, x = input_261)[name = string("linear_183")]; + tensor v_157 = linear(bias = linear_1_bias_0, weight = layers_26_self_attn_v_proj_weight, x = input_261)[name = string("linear_184")]; + tensor var_10469 = const()[name = string("op_10469"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_631 = reshape(shape = var_10469, x = q_209)[name = string("hidden_states_631")]; + tensor var_10475 = const()[name = string("op_10475"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_637 = reshape(shape = var_10475, x = k_209)[name = string("hidden_states_637")]; + tensor var_10481 = const()[name = string("op_10481"), val = tensor([1, 1, 8, 128])]; + tensor v_159 = reshape(shape = var_10481, x = v_157)[name = string("v_159")]; + fp32 var_10486_promoted = const()[name = string("op_10486_promoted"), val = fp32(0x1p+1)]; + tensor var_10492 = pow(x = hidden_states_631, y = var_10486_promoted)[name = string("op_10492")]; + tensor variance_211_axes_0 = const()[name = string("variance_211_axes_0"), val = tensor([-1])]; + bool variance_211_keep_dims_0 = const()[name = string("variance_211_keep_dims_0"), val = bool(true)]; + tensor variance_211 = reduce_mean(axes = variance_211_axes_0, keep_dims = variance_211_keep_dims_0, x = var_10492)[name = string("variance_211")]; + fp32 var_10495 = const()[name = string("op_10495"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10496 = add(x = variance_211, y = var_10495)[name = string("op_10496")]; + fp32 var_10497_epsilon_0 = const()[name = string("op_10497_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10497 = rsqrt(epsilon = var_10497_epsilon_0, x = var_10496)[name = string("op_10497")]; + tensor hidden_states_635 = mul(x = hidden_states_631, y = var_10497)[name = string("hidden_states_635")]; + tensor const_262 = const()[name = string("const_262"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774478784)))]; + tensor q_211 = mul(x = const_262, y = hidden_states_635)[name = string("q_211")]; + fp32 var_10504_promoted = const()[name = string("op_10504_promoted"), val = fp32(0x1p+1)]; + tensor var_10510 = pow(x = hidden_states_637, y = var_10504_promoted)[name = string("op_10510")]; + tensor variance_213_axes_0 = const()[name = string("variance_213_axes_0"), val = tensor([-1])]; + bool variance_213_keep_dims_0 = const()[name = string("variance_213_keep_dims_0"), val = bool(true)]; + tensor variance_213 = reduce_mean(axes = variance_213_axes_0, keep_dims = variance_213_keep_dims_0, x = var_10510)[name = string("variance_213")]; + fp32 var_10513 = const()[name = string("op_10513"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10514 = add(x = variance_213, y = var_10513)[name = string("op_10514")]; + fp32 var_10515_epsilon_0 = const()[name = string("op_10515_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10515 = rsqrt(epsilon = var_10515_epsilon_0, x = var_10514)[name = string("op_10515")]; + tensor hidden_states_641 = mul(x = hidden_states_637, y = var_10515)[name = string("hidden_states_641")]; + tensor const_263 = const()[name = string("const_263"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774479360)))]; + tensor k_211 = mul(x = const_263, y = hidden_states_641)[name = string("k_211")]; + tensor q_213_perm_0 = const()[name = string("q_213_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_213_perm_0 = const()[name = string("k_213_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_161_perm_0 = const()[name = string("v_161_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_213 = transpose(perm = q_213_perm_0, x = q_211)[name = string("transpose_7")]; + tensor var_10532 = mul(x = q_213, y = cos_3)[name = string("op_10532")]; + tensor x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_105 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_213)[name = string("x1_105")]; + tensor x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_105 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_213)[name = string("x2_105")]; + fp32 const_266_promoted = const()[name = string("const_266_promoted"), val = fp32(-0x1p+0)]; + tensor var_10553 = mul(x = x2_105, y = const_266_promoted)[name = string("op_10553")]; + int32 var_10555 = const()[name = string("op_10555"), val = int32(-1)]; + bool var_10556_interleave_0 = const()[name = string("op_10556_interleave_0"), val = bool(false)]; + tensor var_10556 = concat(axis = var_10555, interleave = var_10556_interleave_0, values = (var_10553, x1_105))[name = string("op_10556")]; + tensor var_10557 = mul(x = var_10556, y = sin_3)[name = string("op_10557")]; + tensor q_215 = add(x = var_10532, y = var_10557)[name = string("q_215")]; + tensor k_213 = transpose(perm = k_213_perm_0, x = k_211)[name = string("transpose_6")]; + tensor var_10560 = mul(x = k_213, y = cos_3)[name = string("op_10560")]; + tensor x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_107 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_213)[name = string("x1_107")]; + tensor x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_107 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_213)[name = string("x2_107")]; + fp32 const_269_promoted = const()[name = string("const_269_promoted"), val = fp32(-0x1p+0)]; + tensor var_10581 = mul(x = x2_107, y = const_269_promoted)[name = string("op_10581")]; + int32 var_10583 = const()[name = string("op_10583"), val = int32(-1)]; + bool var_10584_interleave_0 = const()[name = string("op_10584_interleave_0"), val = bool(false)]; + tensor var_10584 = concat(axis = var_10583, interleave = var_10584_interleave_0, values = (var_10581, x1_107))[name = string("op_10584")]; + tensor var_10585 = mul(x = var_10584, y = sin_3)[name = string("op_10585")]; + tensor k_215 = add(x = var_10560, y = var_10585)[name = string("k_215")]; + tensor var_10592 = const()[name = string("op_10592"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat_53 = reshape(shape = var_10592, x = k_215)[name = string("nk_flat_53")]; + tensor var_10598 = const()[name = string("op_10598"), val = tensor([1, 1024, 1, 1])]; + tensor v_161 = transpose(perm = v_161_perm_0, x = v_159)[name = string("transpose_5")]; + tensor nv_flat_53 = reshape(shape = var_10598, x = v_161)[name = string("nv_flat_53")]; + tensor var_10607 = mul(x = var_10401, y = var_1194)[name = string("op_10607")]; + tensor var_10608 = mul(x = nk_flat_53, y = update_mask_1)[name = string("op_10608")]; + tensor key_cache_109 = add(x = var_10607, y = var_10608)[name = string("key_cache_109")]; + tensor var_10614 = mul(x = var_10421, y = var_1194)[name = string("op_10614")]; + tensor var_10615 = mul(x = nv_flat_53, y = update_mask_1)[name = string("op_10615")]; + tensor value_cache_109 = add(x = var_10614, y = var_10615)[name = string("value_cache_109")]; + tensor kc_157_axes_0 = const()[name = string("kc_157_axes_0"), val = tensor([2])]; + tensor kc_157 = squeeze(axes = kc_157_axes_0, x = key_cache_109)[name = string("kc_157")]; + tensor var_10624 = const()[name = string("op_10624"), val = tensor([1, 8, 128, 256])]; + tensor kc_159 = reshape(shape = var_10624, x = kc_157)[name = string("kc_159")]; + tensor vc_157_axes_0 = const()[name = string("vc_157_axes_0"), val = tensor([2])]; + tensor vc_157 = squeeze(axes = vc_157_axes_0, x = value_cache_109)[name = string("vc_157")]; + tensor var_10632 = const()[name = string("op_10632"), val = tensor([1, 8, 128, 256])]; + tensor vc_159 = reshape(shape = var_10632, x = vc_157)[name = string("vc_159")]; + tensor var_10635_axes_0 = const()[name = string("op_10635_axes_0"), val = tensor([2])]; + tensor var_10635 = expand_dims(axes = var_10635_axes_0, x = kc_159)[name = string("op_10635")]; + tensor var_10643_reps_0 = const()[name = string("op_10643_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_10643 = tile(reps = var_10643_reps_0, x = var_10635)[name = string("op_10643")]; + tensor var_10648 = const()[name = string("op_10648"), val = tensor([1, 16, 128, 256])]; + tensor kc_161 = reshape(shape = var_10648, x = var_10643)[name = string("kc_161")]; + tensor var_10651_axes_0 = const()[name = string("op_10651_axes_0"), val = tensor([2])]; + tensor var_10651 = expand_dims(axes = var_10651_axes_0, x = vc_159)[name = string("op_10651")]; + tensor var_10659_reps_0 = const()[name = string("op_10659_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_10659 = tile(reps = var_10659_reps_0, x = var_10651)[name = string("op_10659")]; + tensor var_10664 = const()[name = string("op_10664"), val = tensor([1, 16, 128, 256])]; + tensor vc_161 = reshape(shape = var_10664, x = var_10659)[name = string("vc_161")]; + bool var_10666_transpose_x_0 = const()[name = string("op_10666_transpose_x_0"), val = bool(false)]; + bool var_10666_transpose_y_0 = const()[name = string("op_10666_transpose_y_0"), val = bool(false)]; + tensor var_10666 = matmul(transpose_x = var_10666_transpose_x_0, transpose_y = var_10666_transpose_y_0, x = q_215, y = kc_161)[name = string("op_10666")]; + fp32 _inversed_attn_weights_209_y_0 = const()[name = string("_inversed_attn_weights_209_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_209 = mul(x = var_10666, y = _inversed_attn_weights_209_y_0)[name = string("_inversed_attn_weights_209")]; + tensor attn_weights_211 = add(x = _inversed_attn_weights_209, y = mask_1)[name = string("attn_weights_211")]; + int32 var_10680 = const()[name = string("op_10680"), val = int32(-1)]; + tensor attn_weights_215 = softmax(axis = var_10680, x = attn_weights_211)[name = string("attn_weights_215")]; + bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(false)]; + bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(true)]; + tensor attn_output_105 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_215, y = vc_161)[name = string("attn_output_105")]; + tensor var_10689_perm_0 = const()[name = string("op_10689_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_10693 = const()[name = string("op_10693"), val = tensor([1, 1, -1])]; + tensor var_10689 = transpose(perm = var_10689_perm_0, x = attn_output_105)[name = string("transpose_4")]; + tensor input_263 = reshape(shape = var_10693, x = var_10689)[name = string("input_263")]; + tensor attn_output_107 = linear(bias = linear_1_bias_0, weight = layers_26_self_attn_o_proj_weight, x = input_263)[name = string("linear_185")]; + tensor var_10699_axes_0 = const()[name = string("op_10699_axes_0"), val = tensor([0])]; + tensor var_10699 = squeeze(axes = var_10699_axes_0, x = attn_output_107)[name = string("op_10699")]; + tensor var_10701_axes_0 = const()[name = string("op_10701_axes_0"), val = tensor([0])]; + tensor var_10701 = squeeze(axes = var_10701_axes_0, x = var_10699)[name = string("op_10701")]; + tensor var_10703_axes_0 = const()[name = string("op_10703_axes_0"), val = tensor([-1])]; + tensor var_10703 = expand_dims(axes = var_10703_axes_0, x = var_10701)[name = string("op_10703")]; + tensor attn_4d_53_axes_0 = const()[name = string("attn_4d_53_axes_0"), val = tensor([-1])]; + tensor attn_4d_53 = expand_dims(axes = attn_4d_53_axes_0, x = var_10703)[name = string("attn_4d_53")]; + tensor hidden_105 = add(x = hidden_103, y = attn_4d_53)[name = string("hidden_105")]; + tensor var_10709_axes_0 = const()[name = string("op_10709_axes_0"), val = tensor([-1])]; + tensor var_10709 = squeeze(axes = var_10709_axes_0, x = hidden_105)[name = string("op_10709")]; + tensor var_10711_axes_0 = const()[name = string("op_10711_axes_0"), val = tensor([-1])]; + tensor var_10711 = squeeze(axes = var_10711_axes_0, x = var_10709)[name = string("op_10711")]; + tensor hidden_states_643_axes_0 = const()[name = string("hidden_states_643_axes_0"), val = tensor([0])]; + tensor hidden_states_643 = expand_dims(axes = hidden_states_643_axes_0, x = var_10711)[name = string("hidden_states_643")]; + fp32 var_10717_promoted = const()[name = string("op_10717_promoted"), val = fp32(0x1p+1)]; + tensor var_10723 = pow(x = hidden_states_643, y = var_10717_promoted)[name = string("op_10723")]; + tensor variance_215_axes_0 = const()[name = string("variance_215_axes_0"), val = tensor([-1])]; + bool variance_215_keep_dims_0 = const()[name = string("variance_215_keep_dims_0"), val = bool(true)]; + tensor variance_215 = reduce_mean(axes = variance_215_axes_0, keep_dims = variance_215_keep_dims_0, x = var_10723)[name = string("variance_215")]; + fp32 var_10726 = const()[name = string("op_10726"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10727 = add(x = variance_215, y = var_10726)[name = string("op_10727")]; + fp32 var_10728_epsilon_0 = const()[name = string("op_10728_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10728 = rsqrt(epsilon = var_10728_epsilon_0, x = var_10727)[name = string("op_10728")]; + tensor hidden_states_647 = mul(x = hidden_states_643, y = var_10728)[name = string("hidden_states_647")]; + tensor const_270 = const()[name = string("const_270"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774479936)))]; + tensor input_265 = mul(x = const_270, y = hidden_states_647)[name = string("input_265")]; + tensor input_267 = linear(bias = linear_4_bias_0, weight = layers_26_mlp_gate_proj_weight, x = input_265)[name = string("linear_186")]; + tensor var_10738 = silu(x = input_267)[name = string("op_10738")]; + tensor var_10740 = linear(bias = linear_4_bias_0, weight = layers_26_mlp_up_proj_weight, x = input_265)[name = string("linear_187")]; + tensor input_269 = mul(x = var_10738, y = var_10740)[name = string("input_269")]; + tensor mlp_out_53 = linear(bias = linear_1_bias_0, weight = layers_26_mlp_down_proj_weight, x = input_269)[name = string("linear_188")]; + tensor var_10745_axes_0 = const()[name = string("op_10745_axes_0"), val = tensor([0])]; + tensor var_10745 = squeeze(axes = var_10745_axes_0, x = mlp_out_53)[name = string("op_10745")]; + tensor var_10747_axes_0 = const()[name = string("op_10747_axes_0"), val = tensor([0])]; + tensor var_10747 = squeeze(axes = var_10747_axes_0, x = var_10745)[name = string("op_10747")]; + tensor var_10749_axes_0 = const()[name = string("op_10749_axes_0"), val = tensor([-1])]; + tensor var_10749 = expand_dims(axes = var_10749_axes_0, x = var_10747)[name = string("op_10749")]; + tensor mlp_4d_53_axes_0 = const()[name = string("mlp_4d_53_axes_0"), val = tensor([-1])]; + tensor mlp_4d_53 = expand_dims(axes = mlp_4d_53_axes_0, x = var_10749)[name = string("mlp_4d_53")]; + tensor hidden_107 = add(x = hidden_105, y = mlp_4d_53)[name = string("hidden_107")]; + tensor var_10763_begin_0 = const()[name = string("op_10763_begin_0"), val = tensor([0, 27648, 0, 0])]; + tensor var_10763_end_0 = const()[name = string("op_10763_end_0"), val = tensor([1, 1, 1, 256])]; + tensor var_10763_end_mask_0 = const()[name = string("op_10763_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_10763 = slice_by_index(begin = var_10763_begin_0, end = var_10763_end_0, end_mask = var_10763_end_mask_0, x = cast_3)[name = string("op_10763")]; + tensor var_10783_begin_0 = const()[name = string("op_10783_begin_0"), val = tensor([0, 27648, 0, 0])]; + tensor var_10783_end_0 = const()[name = string("op_10783_end_0"), val = tensor([1, 1, 1, 256])]; + tensor var_10783_end_mask_0 = const()[name = string("op_10783_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_10783 = slice_by_index(begin = var_10783_begin_0, end = var_10783_end_0, end_mask = var_10783_end_mask_0, x = cast_4)[name = string("op_10783")]; + tensor var_10795_axes_0 = const()[name = string("op_10795_axes_0"), val = tensor([-1])]; + tensor var_10795 = squeeze(axes = var_10795_axes_0, x = hidden_107)[name = string("op_10795")]; + tensor var_10797_axes_0 = const()[name = string("op_10797_axes_0"), val = tensor([-1])]; + tensor var_10797 = squeeze(axes = var_10797_axes_0, x = var_10795)[name = string("op_10797")]; + tensor hidden_states_649_axes_0 = const()[name = string("hidden_states_649_axes_0"), val = tensor([0])]; + tensor hidden_states_649 = expand_dims(axes = hidden_states_649_axes_0, x = var_10797)[name = string("hidden_states_649")]; + fp32 var_10803_promoted = const()[name = string("op_10803_promoted"), val = fp32(0x1p+1)]; + tensor var_10809 = pow(x = hidden_states_649, y = var_10803_promoted)[name = string("op_10809")]; + tensor variance_217_axes_0 = const()[name = string("variance_217_axes_0"), val = tensor([-1])]; + bool variance_217_keep_dims_0 = const()[name = string("variance_217_keep_dims_0"), val = bool(true)]; + tensor variance_217 = reduce_mean(axes = variance_217_axes_0, keep_dims = variance_217_keep_dims_0, x = var_10809)[name = string("variance_217")]; + fp32 var_10812 = const()[name = string("op_10812"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10813 = add(x = variance_217, y = var_10812)[name = string("op_10813")]; + fp32 var_10814_epsilon_0 = const()[name = string("op_10814_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10814 = rsqrt(epsilon = var_10814_epsilon_0, x = var_10813)[name = string("op_10814")]; + tensor hidden_states_653 = mul(x = hidden_states_649, y = var_10814)[name = string("hidden_states_653")]; + tensor const_271 = const()[name = string("const_271"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774484096)))]; + tensor input_271 = mul(x = const_271, y = hidden_states_653)[name = string("input_271")]; + tensor q_217 = linear(bias = linear_0_bias_0, weight = layers_27_self_attn_q_proj_weight, x = input_271)[name = string("linear_189")]; + tensor k_217 = linear(bias = linear_1_bias_0, weight = layers_27_self_attn_k_proj_weight, x = input_271)[name = string("linear_190")]; + tensor v_163 = linear(bias = linear_1_bias_0, weight = layers_27_self_attn_v_proj_weight, x = input_271)[name = string("linear_191")]; + tensor var_10831 = const()[name = string("op_10831"), val = tensor([1, 1, 16, 128])]; + tensor hidden_states_655 = reshape(shape = var_10831, x = q_217)[name = string("hidden_states_655")]; + tensor var_10837 = const()[name = string("op_10837"), val = tensor([1, 1, 8, 128])]; + tensor hidden_states_661 = reshape(shape = var_10837, x = k_217)[name = string("hidden_states_661")]; + tensor var_10843 = const()[name = string("op_10843"), val = tensor([1, 1, 8, 128])]; + tensor v_165 = reshape(shape = var_10843, x = v_163)[name = string("v_165")]; + fp32 var_10848_promoted = const()[name = string("op_10848_promoted"), val = fp32(0x1p+1)]; + tensor var_10854 = pow(x = hidden_states_655, y = var_10848_promoted)[name = string("op_10854")]; + tensor variance_219_axes_0 = const()[name = string("variance_219_axes_0"), val = tensor([-1])]; + bool variance_219_keep_dims_0 = const()[name = string("variance_219_keep_dims_0"), val = bool(true)]; + tensor variance_219 = reduce_mean(axes = variance_219_axes_0, keep_dims = variance_219_keep_dims_0, x = var_10854)[name = string("variance_219")]; + fp32 var_10857 = const()[name = string("op_10857"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10858 = add(x = variance_219, y = var_10857)[name = string("op_10858")]; + fp32 var_10859_epsilon_0 = const()[name = string("op_10859_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10859 = rsqrt(epsilon = var_10859_epsilon_0, x = var_10858)[name = string("op_10859")]; + tensor hidden_states_659 = mul(x = hidden_states_655, y = var_10859)[name = string("hidden_states_659")]; + tensor const_272 = const()[name = string("const_272"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774488256)))]; + tensor q_219 = mul(x = const_272, y = hidden_states_659)[name = string("q_219")]; + fp32 var_10866_promoted = const()[name = string("op_10866_promoted"), val = fp32(0x1p+1)]; + tensor var_10872 = pow(x = hidden_states_661, y = var_10866_promoted)[name = string("op_10872")]; + tensor variance_221_axes_0 = const()[name = string("variance_221_axes_0"), val = tensor([-1])]; + bool variance_221_keep_dims_0 = const()[name = string("variance_221_keep_dims_0"), val = bool(true)]; + tensor variance_221 = reduce_mean(axes = variance_221_axes_0, keep_dims = variance_221_keep_dims_0, x = var_10872)[name = string("variance_221")]; + fp32 var_10875 = const()[name = string("op_10875"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_10876 = add(x = variance_221, y = var_10875)[name = string("op_10876")]; + fp32 var_10877_epsilon_0 = const()[name = string("op_10877_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_10877 = rsqrt(epsilon = var_10877_epsilon_0, x = var_10876)[name = string("op_10877")]; + tensor hidden_states_665 = mul(x = hidden_states_661, y = var_10877)[name = string("hidden_states_665")]; + tensor const_273 = const()[name = string("const_273"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774488832)))]; + tensor k_219 = mul(x = const_273, y = hidden_states_665)[name = string("k_219")]; + tensor q_221_perm_0 = const()[name = string("q_221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor k_221_perm_0 = const()[name = string("k_221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor v_perm_0 = const()[name = string("v_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor q_221 = transpose(perm = q_221_perm_0, x = q_219)[name = string("transpose_3")]; + tensor var_10894 = mul(x = q_221, y = cos_3)[name = string("op_10894")]; + tensor x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor([1, 16, 1, 64])]; + tensor x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1_109 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_221)[name = string("x1_109")]; + tensor x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor([1, 16, 1, 128])]; + tensor x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2_109 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_221)[name = string("x2_109")]; + fp32 const_276_promoted = const()[name = string("const_276_promoted"), val = fp32(-0x1p+0)]; + tensor var_10915 = mul(x = x2_109, y = const_276_promoted)[name = string("op_10915")]; + int32 var_10917 = const()[name = string("op_10917"), val = int32(-1)]; + bool var_10918_interleave_0 = const()[name = string("op_10918_interleave_0"), val = bool(false)]; + tensor var_10918 = concat(axis = var_10917, interleave = var_10918_interleave_0, values = (var_10915, x1_109))[name = string("op_10918")]; + tensor var_10919 = mul(x = var_10918, y = sin_3)[name = string("op_10919")]; + tensor q = add(x = var_10894, y = var_10919)[name = string("q")]; + tensor k_221 = transpose(perm = k_221_perm_0, x = k_219)[name = string("transpose_2")]; + tensor var_10922 = mul(x = k_221, y = cos_3)[name = string("op_10922")]; + tensor x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x1_end_0 = const()[name = string("x1_end_0"), val = tensor([1, 8, 1, 64])]; + tensor x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor x1 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_221)[name = string("x1")]; + tensor x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor([0, 0, 0, 64])]; + tensor x2_end_0 = const()[name = string("x2_end_0"), val = tensor([1, 8, 1, 128])]; + tensor x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor([true, true, true, true])]; + tensor x2 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_221)[name = string("x2")]; + fp32 const_279_promoted = const()[name = string("const_279_promoted"), val = fp32(-0x1p+0)]; + tensor var_10943 = mul(x = x2, y = const_279_promoted)[name = string("op_10943")]; + int32 var_10945 = const()[name = string("op_10945"), val = int32(-1)]; + bool var_10946_interleave_0 = const()[name = string("op_10946_interleave_0"), val = bool(false)]; + tensor var_10946 = concat(axis = var_10945, interleave = var_10946_interleave_0, values = (var_10943, x1))[name = string("op_10946")]; + tensor var_10947 = mul(x = var_10946, y = sin_3)[name = string("op_10947")]; + tensor k = add(x = var_10922, y = var_10947)[name = string("k")]; + tensor var_10954 = const()[name = string("op_10954"), val = tensor([1, 1024, 1, 1])]; + tensor nk_flat = reshape(shape = var_10954, x = k)[name = string("nk_flat")]; + tensor var_10960 = const()[name = string("op_10960"), val = tensor([1, 1024, 1, 1])]; + tensor v = transpose(perm = v_perm_0, x = v_165)[name = string("transpose_1")]; + tensor nv_flat = reshape(shape = var_10960, x = v)[name = string("nv_flat")]; + tensor var_10969 = mul(x = var_10763, y = var_1194)[name = string("op_10969")]; + tensor var_10970 = mul(x = nk_flat, y = update_mask_1)[name = string("op_10970")]; + tensor key_cache_1 = add(x = var_10969, y = var_10970)[name = string("key_cache")]; + tensor var_10976 = mul(x = var_10783, y = var_1194)[name = string("op_10976")]; + tensor var_10977 = mul(x = nv_flat, y = update_mask_1)[name = string("op_10977")]; + tensor value_cache_1 = add(x = var_10976, y = var_10977)[name = string("value_cache")]; + tensor kc_163_axes_0 = const()[name = string("kc_163_axes_0"), val = tensor([2])]; + tensor kc_163 = squeeze(axes = kc_163_axes_0, x = key_cache_1)[name = string("kc_163")]; + tensor var_10986 = const()[name = string("op_10986"), val = tensor([1, 8, 128, 256])]; + tensor kc_165 = reshape(shape = var_10986, x = kc_163)[name = string("kc_165")]; + tensor vc_163_axes_0 = const()[name = string("vc_163_axes_0"), val = tensor([2])]; + tensor vc_163 = squeeze(axes = vc_163_axes_0, x = value_cache_1)[name = string("vc_163")]; + tensor var_10994 = const()[name = string("op_10994"), val = tensor([1, 8, 128, 256])]; + tensor vc_165 = reshape(shape = var_10994, x = vc_163)[name = string("vc_165")]; + tensor var_10997_axes_0 = const()[name = string("op_10997_axes_0"), val = tensor([2])]; + tensor var_10997 = expand_dims(axes = var_10997_axes_0, x = kc_165)[name = string("op_10997")]; + tensor var_11005_reps_0 = const()[name = string("op_11005_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_11005 = tile(reps = var_11005_reps_0, x = var_10997)[name = string("op_11005")]; + tensor var_11010 = const()[name = string("op_11010"), val = tensor([1, 16, 128, 256])]; + tensor kc = reshape(shape = var_11010, x = var_11005)[name = string("kc")]; + tensor var_11013_axes_0 = const()[name = string("op_11013_axes_0"), val = tensor([2])]; + tensor var_11013 = expand_dims(axes = var_11013_axes_0, x = vc_165)[name = string("op_11013")]; + tensor var_11021_reps_0 = const()[name = string("op_11021_reps_0"), val = tensor([1, 1, 2, 1, 1])]; + tensor var_11021 = tile(reps = var_11021_reps_0, x = var_11013)[name = string("op_11021")]; + tensor var_11026 = const()[name = string("op_11026"), val = tensor([1, 16, 128, 256])]; + tensor vc = reshape(shape = var_11026, x = var_11021)[name = string("vc")]; + bool var_11028_transpose_x_0 = const()[name = string("op_11028_transpose_x_0"), val = bool(false)]; + bool var_11028_transpose_y_0 = const()[name = string("op_11028_transpose_y_0"), val = bool(false)]; + tensor var_11028 = matmul(transpose_x = var_11028_transpose_x_0, transpose_y = var_11028_transpose_y_0, x = q, y = kc)[name = string("op_11028")]; + fp32 _inversed_attn_weights_217_y_0 = const()[name = string("_inversed_attn_weights_217_y_0"), val = fp32(0x1.6a09e6p-4)]; + tensor _inversed_attn_weights_217 = mul(x = var_11028, y = _inversed_attn_weights_217_y_0)[name = string("_inversed_attn_weights_217")]; + tensor attn_weights_219 = add(x = _inversed_attn_weights_217, y = mask_1)[name = string("attn_weights_219")]; + int32 var_11042 = const()[name = string("op_11042"), val = int32(-1)]; + tensor attn_weights = softmax(axis = var_11042, x = attn_weights_219)[name = string("attn_weights")]; + bool attn_output_109_transpose_x_1 = const()[name = string("attn_output_109_transpose_x_1"), val = bool(false)]; + bool attn_output_109_transpose_y_1 = const()[name = string("attn_output_109_transpose_y_1"), val = bool(true)]; + tensor attn_output_109 = matmul(transpose_x = attn_output_109_transpose_x_1, transpose_y = attn_output_109_transpose_y_1, x = attn_weights, y = vc)[name = string("attn_output_109")]; + tensor var_11051_perm_0 = const()[name = string("op_11051_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_11055 = const()[name = string("op_11055"), val = tensor([1, 1, -1])]; + tensor var_11051 = transpose(perm = var_11051_perm_0, x = attn_output_109)[name = string("transpose_0")]; + tensor input_273 = reshape(shape = var_11055, x = var_11051)[name = string("input_273")]; + tensor attn_output = linear(bias = linear_1_bias_0, weight = layers_27_self_attn_o_proj_weight, x = input_273)[name = string("linear_192")]; + tensor var_11061_axes_0 = const()[name = string("op_11061_axes_0"), val = tensor([0])]; + tensor var_11061 = squeeze(axes = var_11061_axes_0, x = attn_output)[name = string("op_11061")]; + tensor var_11063_axes_0 = const()[name = string("op_11063_axes_0"), val = tensor([0])]; + tensor var_11063 = squeeze(axes = var_11063_axes_0, x = var_11061)[name = string("op_11063")]; + tensor var_11065_axes_0 = const()[name = string("op_11065_axes_0"), val = tensor([-1])]; + tensor var_11065 = expand_dims(axes = var_11065_axes_0, x = var_11063)[name = string("op_11065")]; + tensor attn_4d_axes_0 = const()[name = string("attn_4d_axes_0"), val = tensor([-1])]; + tensor attn_4d = expand_dims(axes = attn_4d_axes_0, x = var_11065)[name = string("attn_4d")]; + tensor hidden_109 = add(x = hidden_107, y = attn_4d)[name = string("hidden_109")]; + tensor var_11071_axes_0 = const()[name = string("op_11071_axes_0"), val = tensor([-1])]; + tensor var_11071 = squeeze(axes = var_11071_axes_0, x = hidden_109)[name = string("op_11071")]; + tensor var_11073_axes_0 = const()[name = string("op_11073_axes_0"), val = tensor([-1])]; + tensor var_11073 = squeeze(axes = var_11073_axes_0, x = var_11071)[name = string("op_11073")]; + tensor hidden_states_667_axes_0 = const()[name = string("hidden_states_667_axes_0"), val = tensor([0])]; + tensor hidden_states_667 = expand_dims(axes = hidden_states_667_axes_0, x = var_11073)[name = string("hidden_states_667")]; + fp32 var_11079_promoted = const()[name = string("op_11079_promoted"), val = fp32(0x1p+1)]; + tensor var_11085 = pow(x = hidden_states_667, y = var_11079_promoted)[name = string("op_11085")]; + tensor variance_223_axes_0 = const()[name = string("variance_223_axes_0"), val = tensor([-1])]; + bool variance_223_keep_dims_0 = const()[name = string("variance_223_keep_dims_0"), val = bool(true)]; + tensor variance_223 = reduce_mean(axes = variance_223_axes_0, keep_dims = variance_223_keep_dims_0, x = var_11085)[name = string("variance_223")]; + fp32 var_11088 = const()[name = string("op_11088"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_11089 = add(x = variance_223, y = var_11088)[name = string("op_11089")]; + fp32 var_11090_epsilon_0 = const()[name = string("op_11090_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_11090 = rsqrt(epsilon = var_11090_epsilon_0, x = var_11089)[name = string("op_11090")]; + tensor hidden_states_671 = mul(x = hidden_states_667, y = var_11090)[name = string("hidden_states_671")]; + tensor const_280 = const()[name = string("const_280"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774489408)))]; + tensor input_275 = mul(x = const_280, y = hidden_states_671)[name = string("input_275")]; + tensor input_277 = linear(bias = linear_4_bias_0, weight = layers_27_mlp_gate_proj_weight, x = input_275)[name = string("linear_193")]; + tensor var_11100 = silu(x = input_277)[name = string("op_11100")]; + tensor var_11102 = linear(bias = linear_4_bias_0, weight = layers_27_mlp_up_proj_weight, x = input_275)[name = string("linear_194")]; + tensor input_279 = mul(x = var_11100, y = var_11102)[name = string("input_279")]; + tensor mlp_out = linear(bias = linear_1_bias_0, weight = layers_27_mlp_down_proj_weight, x = input_279)[name = string("linear_195")]; + tensor var_11107_axes_0 = const()[name = string("op_11107_axes_0"), val = tensor([0])]; + tensor var_11107 = squeeze(axes = var_11107_axes_0, x = mlp_out)[name = string("op_11107")]; + tensor var_11109_axes_0 = const()[name = string("op_11109_axes_0"), val = tensor([0])]; + tensor var_11109 = squeeze(axes = var_11109_axes_0, x = var_11107)[name = string("op_11109")]; + tensor var_11111_axes_0 = const()[name = string("op_11111_axes_0"), val = tensor([-1])]; + tensor var_11111 = expand_dims(axes = var_11111_axes_0, x = var_11109)[name = string("op_11111")]; + tensor mlp_4d_axes_0 = const()[name = string("mlp_4d_axes_0"), val = tensor([-1])]; + tensor mlp_4d = expand_dims(axes = mlp_4d_axes_0, x = var_11111)[name = string("mlp_4d")]; + tensor hidden = add(x = hidden_109, y = mlp_4d)[name = string("hidden")]; + tensor var_11117_axes_0 = const()[name = string("op_11117_axes_0"), val = tensor([-1])]; + tensor var_11117 = squeeze(axes = var_11117_axes_0, x = hidden)[name = string("op_11117")]; + tensor var_11119_axes_0 = const()[name = string("op_11119_axes_0"), val = tensor([-1])]; + tensor var_11119 = squeeze(axes = var_11119_axes_0, x = var_11117)[name = string("op_11119")]; + tensor hidden_states_673_axes_0 = const()[name = string("hidden_states_673_axes_0"), val = tensor([0])]; + tensor hidden_states_673 = expand_dims(axes = hidden_states_673_axes_0, x = var_11119)[name = string("hidden_states_673")]; + fp32 var_11125_promoted = const()[name = string("op_11125_promoted"), val = fp32(0x1p+1)]; + tensor var_11131 = pow(x = hidden_states_673, y = var_11125_promoted)[name = string("op_11131")]; + tensor variance_axes_0 = const()[name = string("variance_axes_0"), val = tensor([-1])]; + bool variance_keep_dims_0 = const()[name = string("variance_keep_dims_0"), val = bool(true)]; + tensor variance = reduce_mean(axes = variance_axes_0, keep_dims = variance_keep_dims_0, x = var_11131)[name = string("variance")]; + fp32 var_11134 = const()[name = string("op_11134"), val = fp32(0x1.0c6f7ap-20)]; + tensor var_11135 = add(x = variance, y = var_11134)[name = string("op_11135")]; + fp32 var_11136_epsilon_0 = const()[name = string("op_11136_epsilon_0"), val = fp32(0x1.197998p-40)]; + tensor var_11136 = rsqrt(epsilon = var_11136_epsilon_0, x = var_11135)[name = string("op_11136")]; + tensor hidden_states_1_1 = mul(x = hidden_states_673, y = var_11136)[name = string("hidden_states")]; + tensor const_281 = const()[name = string("const_281"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1774493568)))]; + tensor input = mul(x = const_281, y = hidden_states_1_1)[name = string("input")]; + tensor logits_type_fp32 = linear(bias = linear_4_bias_0, weight = codec_head_weight, x = input)[name = string("linear_196")]; + int32 var_11144 = const()[name = string("op_11144"), val = int32(1)]; + bool new_kv_1_interleave_0 = const()[name = string("new_kv_1_interleave_0"), val = bool(false)]; + tensor new_kv_1 = concat(axis = var_11144, interleave = new_kv_1_interleave_0, values = (nk_flat_1, nk_flat_3, nk_flat_5, nk_flat_7, nk_flat_9, nk_flat_11, nk_flat_13, nk_flat_15, nk_flat_17, nk_flat_19, nk_flat_21, nk_flat_23, nk_flat_25, nk_flat_27, nk_flat_29, nk_flat_31, nk_flat_33, nk_flat_35, nk_flat_37, nk_flat_39, nk_flat_41, nk_flat_43, nk_flat_45, nk_flat_47, nk_flat_49, nk_flat_51, nk_flat_53, nk_flat))[name = string("new_kv_1")]; + tensor var_11153 = mul(x = cast_3, y = var_1194)[name = string("op_11153")]; + tensor var_11154 = mul(x = new_kv_1, y = update_mask_1)[name = string("op_11154")]; + tensor new_key_cache_type_fp32 = add(x = var_11153, y = var_11154)[name = string("op_11156")]; + int32 var_11158 = const()[name = string("op_11158"), val = int32(1)]; + bool new_kv_interleave_0 = const()[name = string("new_kv_interleave_0"), val = bool(false)]; + tensor new_kv = concat(axis = var_11158, interleave = new_kv_interleave_0, values = (nv_flat_1, nv_flat_3, nv_flat_5, nv_flat_7, nv_flat_9, nv_flat_11, nv_flat_13, nv_flat_15, nv_flat_17, nv_flat_19, nv_flat_21, nv_flat_23, nv_flat_25, nv_flat_27, nv_flat_29, nv_flat_31, nv_flat_33, nv_flat_35, nv_flat_37, nv_flat_39, nv_flat_41, nv_flat_43, nv_flat_45, nv_flat_47, nv_flat_49, nv_flat_51, nv_flat_53, nv_flat))[name = string("new_kv")]; + tensor var_11167 = mul(x = cast_4, y = var_1194)[name = string("op_11167")]; + tensor var_11168 = mul(x = new_kv, y = update_mask_1)[name = string("op_11168")]; + tensor new_value_cache_type_fp32 = add(x = var_11167, y = var_11168)[name = string("op_11170")]; + tensor var_11172_axes_0 = const()[name = string("op_11172_axes_0"), val = tensor([0])]; + tensor var_11172 = squeeze(axes = var_11172_axes_0, x = input)[name = string("op_11172")]; + tensor var_11174_axes_0 = const()[name = string("op_11174_axes_0"), val = tensor([-1])]; + tensor var_11174 = expand_dims(axes = var_11174_axes_0, x = var_11172)[name = string("op_11174")]; + tensor var_11176_axes_0 = const()[name = string("op_11176_axes_0"), val = tensor([-1])]; + tensor hidden_states_type_fp32 = expand_dims(axes = var_11176_axes_0, x = var_11174)[name = string("op_11176")]; + string cast_341_dtype_0 = const()[name = string("cast_341_dtype_0"), val = string("fp16")]; + string cast_342_dtype_0 = const()[name = string("cast_342_dtype_0"), val = string("fp16")]; + string cast_343_dtype_0 = const()[name = string("cast_343_dtype_0"), val = string("fp16")]; + string cast_344_dtype_0 = const()[name = string("cast_344_dtype_0"), val = string("fp16")]; + tensor new_value_cache = cast(dtype = cast_344_dtype_0, x = new_value_cache_type_fp32)[name = string("cast_345")]; + tensor new_key_cache = cast(dtype = cast_343_dtype_0, x = new_key_cache_type_fp32)[name = string("cast_346")]; + tensor hidden_states = cast(dtype = cast_342_dtype_0, x = hidden_states_type_fp32)[name = string("cast_347")]; + tensor logits = cast(dtype = cast_341_dtype_0, x = logits_type_fp32)[name = string("cast_348")]; + } -> (logits, hidden_states, new_key_cache, new_value_cache); +} \ No newline at end of file