diff --git "a/qwen25_FFN_PF_lut6_chunk_03of04.mlmodelc/model.mil" "b/qwen25_FFN_PF_lut6_chunk_03of04.mlmodelc/model.mil"
new file mode 100644--- /dev/null
+++ "b/qwen25_FFN_PF_lut6_chunk_03of04.mlmodelc/model.mil"
@@ -0,0 +1,4047 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func infer<ios18>(tensor<fp16, [1, 1, 1, 2048]> causal_mask, tensor<int32, [1]> current_pos, tensor<fp16, [1, 1, 1536]> hidden_states, state<tensor<fp16, [56, 2, 2048, 128]>> model_model_kv_cache_0, tensor<int32, [1]> position_ids) {
+            tensor<fp16, [1536]> model_model_layers_10_self_attn_q_proj_bias = const()[name = string("model_model_layers_10_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_10_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3200))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1772736))))[name = string("model_model_layers_10_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_10_self_attn_k_proj_bias = const()[name = string("model_model_layers_10_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1821952)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_10_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1822528))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2117504))))[name = string("model_model_layers_10_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_10_self_attn_v_proj_bias = const()[name = string("model_model_layers_10_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2125760)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_10_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2126336))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2421312))))[name = string("model_model_layers_10_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_10_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2429568))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12751552))))[name = string("model_model_layers_10_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_10_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13038336))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23360320))))[name = string("model_model_layers_10_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_10_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23647104))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33969088))))[name = string("model_model_layers_10_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_11_self_attn_q_proj_bias = const()[name = string("model_model_layers_11_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34018304)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_11_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34021440))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35790976))))[name = string("model_model_layers_11_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_11_self_attn_k_proj_bias = const()[name = string("model_model_layers_11_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35840192)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_11_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35840768))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36135744))))[name = string("model_model_layers_11_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_11_self_attn_v_proj_bias = const()[name = string("model_model_layers_11_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36144000)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_11_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36144576))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36439552))))[name = string("model_model_layers_11_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_11_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36447808))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46769792))))[name = string("model_model_layers_11_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_11_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47056576))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57378560))))[name = string("model_model_layers_11_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_11_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57665344))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67987328))))[name = string("model_model_layers_11_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_12_self_attn_q_proj_bias = const()[name = string("model_model_layers_12_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68036544)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_12_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68039680))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69809216))))[name = string("model_model_layers_12_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_12_self_attn_k_proj_bias = const()[name = string("model_model_layers_12_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69858432)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_12_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69859008))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70153984))))[name = string("model_model_layers_12_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_12_self_attn_v_proj_bias = const()[name = string("model_model_layers_12_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70162240)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_12_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70162816))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70457792))))[name = string("model_model_layers_12_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_12_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70466048))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80788032))))[name = string("model_model_layers_12_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_12_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81074816))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91396800))))[name = string("model_model_layers_12_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_12_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91683584))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102005568))))[name = string("model_model_layers_12_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_13_self_attn_q_proj_bias = const()[name = string("model_model_layers_13_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102054784)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_13_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102057920))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103827456))))[name = string("model_model_layers_13_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_13_self_attn_k_proj_bias = const()[name = string("model_model_layers_13_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103876672)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_13_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103877248))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104172224))))[name = string("model_model_layers_13_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_13_self_attn_v_proj_bias = const()[name = string("model_model_layers_13_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104180480)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_13_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104181056))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104476032))))[name = string("model_model_layers_13_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_13_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104484288))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114806272))))[name = string("model_model_layers_13_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_13_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115093056))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125415040))))[name = string("model_model_layers_13_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_13_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125701824))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136023808))))[name = string("model_model_layers_13_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_14_self_attn_q_proj_bias = const()[name = string("model_model_layers_14_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136073024)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_14_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136076160))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137845696))))[name = string("model_model_layers_14_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_14_self_attn_k_proj_bias = const()[name = string("model_model_layers_14_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137894912)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_14_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137895488))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138190464))))[name = string("model_model_layers_14_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_14_self_attn_v_proj_bias = const()[name = string("model_model_layers_14_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138198720)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_14_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138199296))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138494272))))[name = string("model_model_layers_14_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_14_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138502528))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148824512))))[name = string("model_model_layers_14_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_14_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149111296))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159433280))))[name = string("model_model_layers_14_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_14_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159720064))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170042048))))[name = string("model_model_layers_14_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_15_self_attn_q_proj_bias = const()[name = string("model_model_layers_15_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170091264)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_15_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170094400))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171863936))))[name = string("model_model_layers_15_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_15_self_attn_k_proj_bias = const()[name = string("model_model_layers_15_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171913152)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_15_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171913728))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172208704))))[name = string("model_model_layers_15_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_15_self_attn_v_proj_bias = const()[name = string("model_model_layers_15_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172216960)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_15_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172217536))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172512512))))[name = string("model_model_layers_15_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_15_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172520768))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182842752))))[name = string("model_model_layers_15_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_15_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183129536))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193451520))))[name = string("model_model_layers_15_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_15_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193738304))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204060288))))[name = string("model_model_layers_15_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_16_self_attn_q_proj_bias = const()[name = string("model_model_layers_16_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204109504)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_16_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204112640))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205882176))))[name = string("model_model_layers_16_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_16_self_attn_k_proj_bias = const()[name = string("model_model_layers_16_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205931392)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_16_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205931968))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206226944))))[name = string("model_model_layers_16_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_16_self_attn_v_proj_bias = const()[name = string("model_model_layers_16_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206235200)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_16_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206235776))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206530752))))[name = string("model_model_layers_16_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_16_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206539008))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216860992))))[name = string("model_model_layers_16_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_16_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217147776))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227469760))))[name = string("model_model_layers_16_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_16_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227756544))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238078528))))[name = string("model_model_layers_16_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_17_self_attn_q_proj_bias = const()[name = string("model_model_layers_17_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238127744)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_17_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238130880))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239900416))))[name = string("model_model_layers_17_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_17_self_attn_k_proj_bias = const()[name = string("model_model_layers_17_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239949632)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_17_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239950208))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240245184))))[name = string("model_model_layers_17_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_17_self_attn_v_proj_bias = const()[name = string("model_model_layers_17_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240253440)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_17_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240254016))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240548992))))[name = string("model_model_layers_17_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_17_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240557248))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250879232))))[name = string("model_model_layers_17_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_17_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251166016))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261488000))))[name = string("model_model_layers_17_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_17_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261774784))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272096768))))[name = string("model_model_layers_17_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_18_self_attn_q_proj_bias = const()[name = string("model_model_layers_18_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272145984)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_18_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272149120))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273918656))))[name = string("model_model_layers_18_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_18_self_attn_k_proj_bias = const()[name = string("model_model_layers_18_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273967872)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_18_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273968448))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274263424))))[name = string("model_model_layers_18_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_18_self_attn_v_proj_bias = const()[name = string("model_model_layers_18_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274271680)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_18_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274272256))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274567232))))[name = string("model_model_layers_18_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_18_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274575488))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284897472))))[name = string("model_model_layers_18_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_18_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285184256))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295506240))))[name = string("model_model_layers_18_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_18_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295793024))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306115008))))[name = string("model_model_layers_18_mlp_down_proj_weight_palettized")];
+            int32 var_385_batch_dims_0 = const()[name = string("op_385_batch_dims_0"), val = int32(0)];
+            bool var_385_validate_indices_0 = const()[name = string("op_385_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [1, 4096, 128]> var_377_to_fp16 = const()[name = string("op_377_to_fp16"), val = tensor<fp16, [1, 4096, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306164224)))];
+            string current_pos_to_int16_dtype_0 = const()[name = string("current_pos_to_int16_dtype_0"), val = string("int16")];
+            string cast_78_dtype_0 = const()[name = string("cast_78_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1]> current_pos_to_int16 = cast(dtype = current_pos_to_int16_dtype_0, x = current_pos)[name = string("cast_5")];
+            tensor<int32, [1]> cast_78 = cast(dtype = cast_78_dtype_0, x = current_pos_to_int16)[name = string("cast_4")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_78, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(4096)];
+            tensor<int32, [1]> add_0 = add(x = cast_78, y = slice_by_index_0)[name = string("add_0")];
+            tensor<int32, [1]> select_0 = select(a = cast_78, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_3")];
+            tensor<int32, [1]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_2")];
+            tensor<bool, [1]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(4096)];
+            tensor<int32, [1]> add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")];
+            tensor<int32, [1]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 op_385_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("op_385_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1, 128]> op_385_cast_fp16_cast_uint16_cast_uint16 = gather(axis = op_385_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = var_385_batch_dims_0, indices = select_0_1, validate_indices = var_385_validate_indices_0, x = var_377_to_fp16)[name = string("op_385_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 1, 1, -1])];
+            tensor<fp16, [1, 1, 1, 128]> sin_1_cast_fp16 = reshape(shape = var_390, x = op_385_cast_fp16_cast_uint16_cast_uint16)[name = string("sin_1_cast_fp16")];
+            int32 var_400_axis_0 = const()[name = string("op_400_axis_0"), val = int32(1)];
+            int32 var_400_batch_dims_0 = const()[name = string("op_400_batch_dims_0"), val = int32(0)];
+            bool var_400_validate_indices_0 = const()[name = string("op_400_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [1, 4096, 128]> var_392_to_fp16 = const()[name = string("op_392_to_fp16"), val = tensor<fp16, [1, 4096, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307212864)))];
+            string current_pos_to_uint16_dtype_0 = const()[name = string("current_pos_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1]> current_pos_to_uint16 = cast(dtype = current_pos_to_uint16_dtype_0, x = current_pos)[name = string("cast_1")];
+            tensor<fp16, [1, 1, 128]> var_400_cast_fp16_cast_uint16 = gather(axis = var_400_axis_0, batch_dims = var_400_batch_dims_0, indices = current_pos_to_uint16, validate_indices = var_400_validate_indices_0, x = var_392_to_fp16)[name = string("op_400_cast_fp16_cast_uint16")];
+            tensor<int32, [4]> var_405 = const()[name = string("op_405"), val = tensor<int32, [4]>([1, 1, 1, -1])];
+            tensor<fp16, [1, 1, 1, 128]> cos_1_cast_fp16 = reshape(shape = var_405, x = var_400_cast_fp16_cast_uint16)[name = string("cos_1_cast_fp16")];
+            int32 var_426 = const()[name = string("op_426"), val = int32(-1)];
+            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_428_cast_fp16 = mul(x = hidden_states, y = const_0_promoted_to_fp16)[name = string("op_428_cast_fp16")];
+            bool input_1_interleave_0 = const()[name = string("input_1_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_1_cast_fp16 = concat(axis = var_426, interleave = input_1_interleave_0, values = (hidden_states, var_428_cast_fp16))[name = string("input_1_cast_fp16")];
+            tensor<int32, [1]> normed_1_axes_0 = const()[name = string("normed_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_423_to_fp16 = const()[name = string("op_423_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_1_cast_fp16 = layer_norm(axes = normed_1_axes_0, epsilon = var_423_to_fp16, x = input_1_cast_fp16)[name = string("normed_1_cast_fp16")];
+            tensor<int32, [3]> normed_3_begin_0 = const()[name = string("normed_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_3_end_0 = const()[name = string("normed_3_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_3_end_mask_0 = const()[name = string("normed_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_3_cast_fp16 = slice_by_index(begin = normed_3_begin_0, end = normed_3_end_0, end_mask = normed_3_end_mask_0, x = normed_1_cast_fp16)[name = string("normed_3_cast_fp16")];
+            tensor<fp16, [1536]> const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308261504)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_3_cast_fp16 = mul(x = normed_3_cast_fp16, y = const_3_promoted_to_fp16)[name = string("hidden_states_3_cast_fp16")];
+            tensor<int32, [3]> var_445 = const()[name = string("op_445"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_448_axes_0 = const()[name = string("op_448_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_446_cast_fp16 = transpose(perm = var_445, x = hidden_states_3_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [1, 1536, 1, 1]> var_448_cast_fp16 = expand_dims(axes = var_448_axes_0, x = var_446_cast_fp16)[name = string("op_448_cast_fp16")];
+            string var_464_pad_type_0 = const()[name = string("op_464_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_464_strides_0 = const()[name = string("op_464_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_464_pad_0 = const()[name = string("op_464_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_464_dilations_0 = const()[name = string("op_464_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_464_groups_0 = const()[name = string("op_464_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_464 = conv(bias = model_model_layers_10_self_attn_q_proj_bias, dilations = var_464_dilations_0, groups = var_464_groups_0, pad = var_464_pad_0, pad_type = var_464_pad_type_0, strides = var_464_strides_0, weight = model_model_layers_10_self_attn_q_proj_weight_palettized, x = var_448_cast_fp16)[name = string("op_464")];
+            tensor<int32, [4]> var_469 = const()[name = string("op_469"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_470 = reshape(shape = var_469, x = var_464)[name = string("op_470")];
+            string var_486_pad_type_0 = const()[name = string("op_486_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_486_strides_0 = const()[name = string("op_486_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_486_pad_0 = const()[name = string("op_486_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_486_dilations_0 = const()[name = string("op_486_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_486_groups_0 = const()[name = string("op_486_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_486 = conv(bias = model_model_layers_10_self_attn_k_proj_bias, dilations = var_486_dilations_0, groups = var_486_groups_0, pad = var_486_pad_0, pad_type = var_486_pad_type_0, strides = var_486_strides_0, weight = model_model_layers_10_self_attn_k_proj_weight_palettized, x = var_448_cast_fp16)[name = string("op_486")];
+            tensor<int32, [4]> var_491 = const()[name = string("op_491"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_492 = reshape(shape = var_491, x = var_486)[name = string("op_492")];
+            string var_508_pad_type_0 = const()[name = string("op_508_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_508_strides_0 = const()[name = string("op_508_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_508_pad_0 = const()[name = string("op_508_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_508_dilations_0 = const()[name = string("op_508_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_508_groups_0 = const()[name = string("op_508_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_508 = conv(bias = model_model_layers_10_self_attn_v_proj_bias, dilations = var_508_dilations_0, groups = var_508_groups_0, pad = var_508_pad_0, pad_type = var_508_pad_type_0, strides = var_508_strides_0, weight = model_model_layers_10_self_attn_v_proj_weight_palettized, x = var_448_cast_fp16)[name = string("op_508")];
+            tensor<int32, [4]> var_513 = const()[name = string("op_513"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_514 = reshape(shape = var_513, x = var_508)[name = string("op_514")];
+            tensor<fp16, [1, 12, 1, 128]> var_520 = mul(x = var_470, y = cos_1_cast_fp16)[name = string("op_520")];
+            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_1 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = var_470)[name = string("x1_1")];
+            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_1 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = var_470)[name = string("x2_1")];
+            fp16 const_6_promoted = const()[name = string("const_6_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_541 = mul(x = x2_1, y = const_6_promoted)[name = string("op_541")];
+            int32 var_543 = const()[name = string("op_543"), val = int32(-1)];
+            bool var_544_interleave_0 = const()[name = string("op_544_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_544 = concat(axis = var_543, interleave = var_544_interleave_0, values = (var_541, x1_1))[name = string("op_544")];
+            tensor<fp16, [1, 12, 1, 128]> var_545 = mul(x = var_544, y = sin_1_cast_fp16)[name = string("op_545")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_1 = add(x = var_520, y = var_545)[name = string("query_states_1")];
+            tensor<fp16, [1, 2, 1, 128]> var_548 = mul(x = var_492, y = cos_1_cast_fp16)[name = string("op_548")];
+            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_3 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = var_492)[name = string("x1_3")];
+            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_3 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = var_492)[name = string("x2_3")];
+            fp16 const_9_promoted = const()[name = string("const_9_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_569 = mul(x = x2_3, y = const_9_promoted)[name = string("op_569")];
+            int32 var_571 = const()[name = string("op_571"), val = int32(-1)];
+            bool var_572_interleave_0 = const()[name = string("op_572_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_572 = concat(axis = var_571, interleave = var_572_interleave_0, values = (var_569, x1_3))[name = string("op_572")];
+            tensor<fp16, [1, 2, 1, 128]> var_573 = mul(x = var_572, y = sin_1_cast_fp16)[name = string("op_573")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_1 = add(x = var_548, y = var_573)[name = string("key_states_1")];
+            int32 var_577 = const()[name = string("op_577"), val = int32(1)];
+            tensor<int32, [1]> var_578 = add(x = current_pos, y = var_577)[name = string("op_578")];
+            tensor<fp16, [56, 2, 2048, 128]> read_state_0 = read_state(input = model_model_kv_cache_0)[name = string("read_state_0")];
+            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([10])];
+            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([11])];
+            int32 concat_2_axis_0 = const()[name = string("concat_2_axis_0"), val = int32(0)];
+            bool concat_2_interleave_0 = const()[name = string("concat_2_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_2 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = (expand_dims_0, expand_dims_1, current_pos, expand_dims_3))[name = string("concat_2")];
+            tensor<int32, [1]> concat_3_values1_0 = const()[name = string("concat_3_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_3_values3_0 = const()[name = string("concat_3_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_3_axis_0 = const()[name = string("concat_3_axis_0"), val = int32(0)];
+            bool concat_3_interleave_0 = const()[name = string("concat_3_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (expand_dims_4, concat_3_values1_0, var_578, concat_3_values3_0))[name = string("concat_3")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_1_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_2, begin_mask = model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_1_stride_0, update = key_states_1, x = read_state_0)[name = string("model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_0_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_18 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_0")];
+            tensor<int32, [1]> expand_dims_6 = const()[name = string("expand_dims_6"), val = tensor<int32, [1]>([38])];
+            tensor<int32, [1]> expand_dims_7 = const()[name = string("expand_dims_7"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_9 = const()[name = string("expand_dims_9"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_10 = const()[name = string("expand_dims_10"), val = tensor<int32, [1]>([39])];
+            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
+            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_6, expand_dims_7, current_pos, expand_dims_9))[name = string("concat_6")];
+            tensor<int32, [1]> concat_7_values1_0 = const()[name = string("concat_7_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_7_values3_0 = const()[name = string("concat_7_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_7_axis_0 = const()[name = string("concat_7_axis_0"), val = int32(0)];
+            bool concat_7_interleave_0 = const()[name = string("concat_7_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_7 = concat(axis = concat_7_axis_0, interleave = concat_7_interleave_0, values = (expand_dims_10, concat_7_values1_0, var_578, concat_7_values3_0))[name = string("concat_7")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_2_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_6, begin_mask = model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0, end = concat_7, end_mask = model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_2_stride_0, update = var_514, x = coreml_update_state_18)[name = string("model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_1_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_19 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_1")];
+            tensor<int32, [4]> var_628_begin_0 = const()[name = string("op_628_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> var_628_end_0 = const()[name = string("op_628_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
+            tensor<bool, [4]> var_628_end_mask_0 = const()[name = string("op_628_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_628_cast_fp16 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = coreml_update_state_19)[name = string("op_628_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_1_axes_0 = const()[name = string("K_layer_cache_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_1_cast_fp16 = squeeze(axes = K_layer_cache_1_axes_0, x = var_628_cast_fp16)[name = string("K_layer_cache_1_cast_fp16")];
+            tensor<int32, [4]> var_635_begin_0 = const()[name = string("op_635_begin_0"), val = tensor<int32, [4]>([38, 0, 0, 0])];
+            tensor<int32, [4]> var_635_end_0 = const()[name = string("op_635_end_0"), val = tensor<int32, [4]>([39, 2, 2048, 128])];
+            tensor<bool, [4]> var_635_end_mask_0 = const()[name = string("op_635_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_635_cast_fp16 = slice_by_index(begin = var_635_begin_0, end = var_635_end_0, end_mask = var_635_end_mask_0, x = coreml_update_state_19)[name = string("op_635_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_1_axes_0 = const()[name = string("V_layer_cache_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_1_cast_fp16 = squeeze(axes = V_layer_cache_1_axes_0, x = var_635_cast_fp16)[name = string("V_layer_cache_1_cast_fp16")];
+            tensor<int32, [1]> x_3_axes_0 = const()[name = string("x_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_3_cast_fp16 = expand_dims(axes = x_3_axes_0, x = K_layer_cache_1_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_672 = const()[name = string("op_672"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_5_cast_fp16 = tile(reps = var_672, x = x_3_cast_fp16)[name = string("x_5_cast_fp16")];
+            tensor<int32, [4]> var_684 = const()[name = string("op_684"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_3_cast_fp16 = reshape(shape = var_684, x = x_5_cast_fp16)[name = string("key_states_3_cast_fp16")];
+            tensor<int32, [1]> x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_9_cast_fp16 = expand_dims(axes = x_9_axes_0, x = V_layer_cache_1_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_11_cast_fp16 = tile(reps = var_692, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<int32, [4]> var_704 = const()[name = string("op_704"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_3_cast_fp16 = reshape(shape = var_704, x = x_11_cast_fp16)[name = string("value_states_3_cast_fp16")];
+            bool var_727_transpose_x_1 = const()[name = string("op_727_transpose_x_1"), val = bool(false)];
+            bool var_727_transpose_y_1 = const()[name = string("op_727_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_727_cast_fp16 = matmul(transpose_x = var_727_transpose_x_1, transpose_y = var_727_transpose_y_1, x = query_states_1, y = key_states_3_cast_fp16)[name = string("op_727_cast_fp16")];
+            fp16 var_728_to_fp16 = const()[name = string("op_728_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_1_cast_fp16 = mul(x = var_727_cast_fp16, y = var_728_to_fp16)[name = string("attn_logits_1_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_3_cast_fp16 = add(x = attn_logits_1_cast_fp16, y = causal_mask)[name = string("attn_logits_3_cast_fp16")];
+            int32 var_755 = const()[name = string("op_755"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_757_cast_fp16 = softmax(axis = var_755, x = attn_logits_3_cast_fp16)[name = string("op_757_cast_fp16")];
+            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
+            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_757_cast_fp16, y = value_states_3_cast_fp16)[name = string("attn_output_1_cast_fp16")];
+            tensor<int32, [4]> var_781_perm_0 = const()[name = string("op_781_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_785 = const()[name = string("op_785"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_781 = transpose(perm = var_781_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [1, 1, 1536]> attn_output_7 = reshape(shape = var_785, x = var_781)[name = string("attn_output_7")];
+            tensor<int32, [3]> var_790 = const()[name = string("op_790"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_0_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308264640))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310034176))))[name = string("squeeze_0_palettized")];
+            string var_806_pad_type_0 = const()[name = string("op_806_pad_type_0"), val = string("valid")];
+            int32 var_806_groups_0 = const()[name = string("op_806_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_806_strides_0 = const()[name = string("op_806_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_806_pad_0 = const()[name = string("op_806_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_806_dilations_0 = const()[name = string("op_806_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_791 = transpose(perm = var_790, x = attn_output_7)[name = string("transpose_51")];
+            tensor<fp16, [1, 1536, 1]> var_806 = conv(dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = squeeze_0_palettized, x = var_791)[name = string("op_806")];
+            tensor<int32, [3]> var_810 = const()[name = string("op_810"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_11 = transpose(perm = var_810, x = var_806)[name = string("transpose_50")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_5_cast_fp16 = add(x = hidden_states, y = attn_output_11)[name = string("hidden_states_5_cast_fp16")];
+            int32 var_823 = const()[name = string("op_823"), val = int32(-1)];
+            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_825_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_825_cast_fp16")];
+            bool input_7_interleave_0 = const()[name = string("input_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_7_cast_fp16 = concat(axis = var_823, interleave = input_7_interleave_0, values = (hidden_states_5_cast_fp16, var_825_cast_fp16))[name = string("input_7_cast_fp16")];
+            tensor<int32, [1]> normed_5_axes_0 = const()[name = string("normed_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_820_to_fp16 = const()[name = string("op_820_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_5_cast_fp16 = layer_norm(axes = normed_5_axes_0, epsilon = var_820_to_fp16, x = input_7_cast_fp16)[name = string("normed_5_cast_fp16")];
+            tensor<int32, [3]> normed_7_begin_0 = const()[name = string("normed_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_7_end_0 = const()[name = string("normed_7_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_7_end_mask_0 = const()[name = string("normed_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_7_cast_fp16 = slice_by_index(begin = normed_7_begin_0, end = normed_7_end_0, end_mask = normed_7_end_mask_0, x = normed_5_cast_fp16)[name = string("normed_7_cast_fp16")];
+            tensor<fp16, [1536]> const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310083392)))];
+            tensor<fp16, [1, 1, 1536]> x_13_cast_fp16 = mul(x = normed_7_cast_fp16, y = const_21_promoted_to_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [3]> var_850 = const()[name = string("op_850"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_851 = transpose(perm = var_850, x = x_13_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 1536, 1, 1]> input_9 = expand_dims(axes = input_9_axes_0, x = var_851)[name = string("input_9")];
+            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_11 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = model_model_layers_10_mlp_gate_proj_weight_palettized, x = input_9)[name = string("input_11")];
+            string b_1_pad_type_0 = const()[name = string("b_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_1_strides_0 = const()[name = string("b_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_1_pad_0 = const()[name = string("b_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_1_dilations_0 = const()[name = string("b_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_1_groups_0 = const()[name = string("b_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_1 = conv(dilations = b_1_dilations_0, groups = b_1_groups_0, pad = b_1_pad_0, pad_type = b_1_pad_type_0, strides = b_1_strides_0, weight = model_model_layers_10_mlp_up_proj_weight_palettized, x = input_9)[name = string("b_1")];
+            tensor<fp16, [1, 8960, 1, 1]> c_1 = silu(x = input_11)[name = string("c_1")];
+            tensor<fp16, [1, 8960, 1, 1]> input_13 = mul(x = c_1, y = b_1)[name = string("input_13")];
+            string e_1_pad_type_0 = const()[name = string("e_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_1_strides_0 = const()[name = string("e_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_1_pad_0 = const()[name = string("e_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_1_dilations_0 = const()[name = string("e_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_1_groups_0 = const()[name = string("e_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_1 = conv(dilations = e_1_dilations_0, groups = e_1_groups_0, pad = e_1_pad_0, pad_type = e_1_pad_type_0, strides = e_1_strides_0, weight = model_model_layers_10_mlp_down_proj_weight_palettized, x = input_13)[name = string("e_1")];
+            tensor<int32, [1]> var_873_axes_0 = const()[name = string("op_873_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_873 = squeeze(axes = var_873_axes_0, x = e_1)[name = string("op_873")];
+            tensor<int32, [3]> var_874 = const()[name = string("op_874"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_875 = transpose(perm = var_874, x = var_873)[name = string("transpose_48")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_7_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = var_875)[name = string("hidden_states_7_cast_fp16")];
+            int32 var_887 = const()[name = string("op_887"), val = int32(-1)];
+            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_889_cast_fp16 = mul(x = hidden_states_7_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_889_cast_fp16")];
+            bool input_15_interleave_0 = const()[name = string("input_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_15_cast_fp16 = concat(axis = var_887, interleave = input_15_interleave_0, values = (hidden_states_7_cast_fp16, var_889_cast_fp16))[name = string("input_15_cast_fp16")];
+            tensor<int32, [1]> normed_9_axes_0 = const()[name = string("normed_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_884_to_fp16 = const()[name = string("op_884_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_9_cast_fp16 = layer_norm(axes = normed_9_axes_0, epsilon = var_884_to_fp16, x = input_15_cast_fp16)[name = string("normed_9_cast_fp16")];
+            tensor<int32, [3]> normed_11_begin_0 = const()[name = string("normed_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_11_end_0 = const()[name = string("normed_11_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_11_end_mask_0 = const()[name = string("normed_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_11_cast_fp16 = slice_by_index(begin = normed_11_begin_0, end = normed_11_end_0, end_mask = normed_11_end_mask_0, x = normed_9_cast_fp16)[name = string("normed_11_cast_fp16")];
+            tensor<fp16, [1536]> const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310086528)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_9_cast_fp16 = mul(x = normed_11_cast_fp16, y = const_25_promoted_to_fp16)[name = string("hidden_states_9_cast_fp16")];
+            tensor<int32, [3]> var_906 = const()[name = string("op_906"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_909_axes_0 = const()[name = string("op_909_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_907_cast_fp16 = transpose(perm = var_906, x = hidden_states_9_cast_fp16)[name = string("transpose_47")];
+            tensor<fp16, [1, 1536, 1, 1]> var_909_cast_fp16 = expand_dims(axes = var_909_axes_0, x = var_907_cast_fp16)[name = string("op_909_cast_fp16")];
+            string var_925_pad_type_0 = const()[name = string("op_925_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_925_strides_0 = const()[name = string("op_925_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_925_pad_0 = const()[name = string("op_925_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_925_dilations_0 = const()[name = string("op_925_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_925_groups_0 = const()[name = string("op_925_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_925 = conv(bias = model_model_layers_11_self_attn_q_proj_bias, dilations = var_925_dilations_0, groups = var_925_groups_0, pad = var_925_pad_0, pad_type = var_925_pad_type_0, strides = var_925_strides_0, weight = model_model_layers_11_self_attn_q_proj_weight_palettized, x = var_909_cast_fp16)[name = string("op_925")];
+            tensor<int32, [4]> var_930 = const()[name = string("op_930"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_931 = reshape(shape = var_930, x = var_925)[name = string("op_931")];
+            string var_947_pad_type_0 = const()[name = string("op_947_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_947_strides_0 = const()[name = string("op_947_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_947_pad_0 = const()[name = string("op_947_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_947_dilations_0 = const()[name = string("op_947_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_947_groups_0 = const()[name = string("op_947_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_947 = conv(bias = model_model_layers_11_self_attn_k_proj_bias, dilations = var_947_dilations_0, groups = var_947_groups_0, pad = var_947_pad_0, pad_type = var_947_pad_type_0, strides = var_947_strides_0, weight = model_model_layers_11_self_attn_k_proj_weight_palettized, x = var_909_cast_fp16)[name = string("op_947")];
+            tensor<int32, [4]> var_952 = const()[name = string("op_952"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_953 = reshape(shape = var_952, x = var_947)[name = string("op_953")];
+            string var_969_pad_type_0 = const()[name = string("op_969_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_969_strides_0 = const()[name = string("op_969_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_969_pad_0 = const()[name = string("op_969_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_969_dilations_0 = const()[name = string("op_969_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_969_groups_0 = const()[name = string("op_969_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_969 = conv(bias = model_model_layers_11_self_attn_v_proj_bias, dilations = var_969_dilations_0, groups = var_969_groups_0, pad = var_969_pad_0, pad_type = var_969_pad_type_0, strides = var_969_strides_0, weight = model_model_layers_11_self_attn_v_proj_weight_palettized, x = var_909_cast_fp16)[name = string("op_969")];
+            tensor<int32, [4]> var_974 = const()[name = string("op_974"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_975 = reshape(shape = var_974, x = var_969)[name = string("op_975")];
+            tensor<fp16, [1, 12, 1, 128]> var_981 = mul(x = var_931, y = cos_1_cast_fp16)[name = string("op_981")];
+            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_5 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = var_931)[name = string("x1_5")];
+            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_5 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = var_931)[name = string("x2_5")];
+            fp16 const_28_promoted = const()[name = string("const_28_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_1002 = mul(x = x2_5, y = const_28_promoted)[name = string("op_1002")];
+            int32 var_1004 = const()[name = string("op_1004"), val = int32(-1)];
+            bool var_1005_interleave_0 = const()[name = string("op_1005_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_1005 = concat(axis = var_1004, interleave = var_1005_interleave_0, values = (var_1002, x1_5))[name = string("op_1005")];
+            tensor<fp16, [1, 12, 1, 128]> var_1006 = mul(x = var_1005, y = sin_1_cast_fp16)[name = string("op_1006")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_3 = add(x = var_981, y = var_1006)[name = string("query_states_3")];
+            tensor<fp16, [1, 2, 1, 128]> var_1009 = mul(x = var_953, y = cos_1_cast_fp16)[name = string("op_1009")];
+            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_7 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = var_953)[name = string("x1_7")];
+            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_7 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = var_953)[name = string("x2_7")];
+            fp16 const_31_promoted = const()[name = string("const_31_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_1030 = mul(x = x2_7, y = const_31_promoted)[name = string("op_1030")];
+            int32 var_1032 = const()[name = string("op_1032"), val = int32(-1)];
+            bool var_1033_interleave_0 = const()[name = string("op_1033_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_1033 = concat(axis = var_1032, interleave = var_1033_interleave_0, values = (var_1030, x1_7))[name = string("op_1033")];
+            tensor<fp16, [1, 2, 1, 128]> var_1034 = mul(x = var_1033, y = sin_1_cast_fp16)[name = string("op_1034")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_5 = add(x = var_1009, y = var_1034)[name = string("key_states_5")];
+            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([11])];
+            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([12])];
+            int32 concat_10_axis_0 = const()[name = string("concat_10_axis_0"), val = int32(0)];
+            bool concat_10_interleave_0 = const()[name = string("concat_10_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_10 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (expand_dims_12, expand_dims_13, current_pos, expand_dims_15))[name = string("concat_10")];
+            tensor<int32, [1]> concat_11_values1_0 = const()[name = string("concat_11_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_11_values3_0 = const()[name = string("concat_11_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_11_axis_0 = const()[name = string("concat_11_axis_0"), val = int32(0)];
+            bool concat_11_interleave_0 = const()[name = string("concat_11_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (expand_dims_16, concat_11_values1_0, var_578, concat_11_values3_0))[name = string("concat_11")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_3_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_10, begin_mask = model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0, end = concat_11, end_mask = model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_3_stride_0, update = key_states_5, x = coreml_update_state_19)[name = string("model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_2_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_20 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_2")];
+            tensor<int32, [1]> expand_dims_18 = const()[name = string("expand_dims_18"), val = tensor<int32, [1]>([39])];
+            tensor<int32, [1]> expand_dims_19 = const()[name = string("expand_dims_19"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_21 = const()[name = string("expand_dims_21"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_22 = const()[name = string("expand_dims_22"), val = tensor<int32, [1]>([40])];
+            int32 concat_14_axis_0 = const()[name = string("concat_14_axis_0"), val = int32(0)];
+            bool concat_14_interleave_0 = const()[name = string("concat_14_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_14 = concat(axis = concat_14_axis_0, interleave = concat_14_interleave_0, values = (expand_dims_18, expand_dims_19, current_pos, expand_dims_21))[name = string("concat_14")];
+            tensor<int32, [1]> concat_15_values1_0 = const()[name = string("concat_15_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_15_values3_0 = const()[name = string("concat_15_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_15_axis_0 = const()[name = string("concat_15_axis_0"), val = int32(0)];
+            bool concat_15_interleave_0 = const()[name = string("concat_15_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (expand_dims_22, concat_15_values1_0, var_578, concat_15_values3_0))[name = string("concat_15")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_4_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_14, begin_mask = model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0, end = concat_15, end_mask = model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_4_stride_0, update = var_975, x = coreml_update_state_20)[name = string("model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_3_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_21 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_3")];
+            tensor<int32, [4]> var_1089_begin_0 = const()[name = string("op_1089_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> var_1089_end_0 = const()[name = string("op_1089_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
+            tensor<bool, [4]> var_1089_end_mask_0 = const()[name = string("op_1089_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1089_cast_fp16 = slice_by_index(begin = var_1089_begin_0, end = var_1089_end_0, end_mask = var_1089_end_mask_0, x = coreml_update_state_21)[name = string("op_1089_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_3_axes_0 = const()[name = string("K_layer_cache_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_3_cast_fp16 = squeeze(axes = K_layer_cache_3_axes_0, x = var_1089_cast_fp16)[name = string("K_layer_cache_3_cast_fp16")];
+            tensor<int32, [4]> var_1096_begin_0 = const()[name = string("op_1096_begin_0"), val = tensor<int32, [4]>([39, 0, 0, 0])];
+            tensor<int32, [4]> var_1096_end_0 = const()[name = string("op_1096_end_0"), val = tensor<int32, [4]>([40, 2, 2048, 128])];
+            tensor<bool, [4]> var_1096_end_mask_0 = const()[name = string("op_1096_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1096_cast_fp16 = slice_by_index(begin = var_1096_begin_0, end = var_1096_end_0, end_mask = var_1096_end_mask_0, x = coreml_update_state_21)[name = string("op_1096_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_3_axes_0 = const()[name = string("V_layer_cache_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_3_cast_fp16 = squeeze(axes = V_layer_cache_3_axes_0, x = var_1096_cast_fp16)[name = string("V_layer_cache_3_cast_fp16")];
+            tensor<int32, [1]> x_19_axes_0 = const()[name = string("x_19_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_19_cast_fp16 = expand_dims(axes = x_19_axes_0, x = K_layer_cache_3_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [4]> var_1133 = const()[name = string("op_1133"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_21_cast_fp16 = tile(reps = var_1133, x = x_19_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_7_cast_fp16 = reshape(shape = var_1145, x = x_21_cast_fp16)[name = string("key_states_7_cast_fp16")];
+            tensor<int32, [1]> x_25_axes_0 = const()[name = string("x_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_25_cast_fp16 = expand_dims(axes = x_25_axes_0, x = V_layer_cache_3_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [4]> var_1153 = const()[name = string("op_1153"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_27_cast_fp16 = tile(reps = var_1153, x = x_25_cast_fp16)[name = string("x_27_cast_fp16")];
+            tensor<int32, [4]> var_1165 = const()[name = string("op_1165"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_7_cast_fp16 = reshape(shape = var_1165, x = x_27_cast_fp16)[name = string("value_states_7_cast_fp16")];
+            bool var_1188_transpose_x_1 = const()[name = string("op_1188_transpose_x_1"), val = bool(false)];
+            bool var_1188_transpose_y_1 = const()[name = string("op_1188_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_1188_cast_fp16 = matmul(transpose_x = var_1188_transpose_x_1, transpose_y = var_1188_transpose_y_1, x = query_states_3, y = key_states_7_cast_fp16)[name = string("op_1188_cast_fp16")];
+            fp16 var_1189_to_fp16 = const()[name = string("op_1189_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_5_cast_fp16 = mul(x = var_1188_cast_fp16, y = var_1189_to_fp16)[name = string("attn_logits_5_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_7_cast_fp16 = add(x = attn_logits_5_cast_fp16, y = causal_mask)[name = string("attn_logits_7_cast_fp16")];
+            int32 var_1216 = const()[name = string("op_1216"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_1218_cast_fp16 = softmax(axis = var_1216, x = attn_logits_7_cast_fp16)[name = string("op_1218_cast_fp16")];
+            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
+            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_1218_cast_fp16, y = value_states_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
+            tensor<int32, [4]> var_1242_perm_0 = const()[name = string("op_1242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1246 = const()[name = string("op_1246"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_1242 = transpose(perm = var_1242_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [1, 1, 1536]> attn_output_19 = reshape(shape = var_1246, x = var_1242)[name = string("attn_output_19")];
+            tensor<int32, [3]> var_1251 = const()[name = string("op_1251"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_1_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310089664))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311859200))))[name = string("squeeze_1_palettized")];
+            string var_1267_pad_type_0 = const()[name = string("op_1267_pad_type_0"), val = string("valid")];
+            int32 var_1267_groups_0 = const()[name = string("op_1267_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_1267_strides_0 = const()[name = string("op_1267_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_1267_pad_0 = const()[name = string("op_1267_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_1267_dilations_0 = const()[name = string("op_1267_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_1252 = transpose(perm = var_1251, x = attn_output_19)[name = string("transpose_45")];
+            tensor<fp16, [1, 1536, 1]> var_1267 = conv(dilations = var_1267_dilations_0, groups = var_1267_groups_0, pad = var_1267_pad_0, pad_type = var_1267_pad_type_0, strides = var_1267_strides_0, weight = squeeze_1_palettized, x = var_1252)[name = string("op_1267")];
+            tensor<int32, [3]> var_1271 = const()[name = string("op_1271"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_23 = transpose(perm = var_1271, x = var_1267)[name = string("transpose_44")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_11_cast_fp16 = add(x = hidden_states_7_cast_fp16, y = attn_output_23)[name = string("hidden_states_11_cast_fp16")];
+            int32 var_1284 = const()[name = string("op_1284"), val = int32(-1)];
+            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1286_cast_fp16 = mul(x = hidden_states_11_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1286_cast_fp16")];
+            bool input_21_interleave_0 = const()[name = string("input_21_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_21_cast_fp16 = concat(axis = var_1284, interleave = input_21_interleave_0, values = (hidden_states_11_cast_fp16, var_1286_cast_fp16))[name = string("input_21_cast_fp16")];
+            tensor<int32, [1]> normed_13_axes_0 = const()[name = string("normed_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1281_to_fp16 = const()[name = string("op_1281_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_13_cast_fp16 = layer_norm(axes = normed_13_axes_0, epsilon = var_1281_to_fp16, x = input_21_cast_fp16)[name = string("normed_13_cast_fp16")];
+            tensor<int32, [3]> normed_15_begin_0 = const()[name = string("normed_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_15_end_0 = const()[name = string("normed_15_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_15_end_mask_0 = const()[name = string("normed_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_15_cast_fp16 = slice_by_index(begin = normed_15_begin_0, end = normed_15_end_0, end_mask = normed_15_end_mask_0, x = normed_13_cast_fp16)[name = string("normed_15_cast_fp16")];
+            tensor<fp16, [1536]> const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311908416)))];
+            tensor<fp16, [1, 1, 1536]> x_29_cast_fp16 = mul(x = normed_15_cast_fp16, y = const_43_promoted_to_fp16)[name = string("x_29_cast_fp16")];
+            tensor<int32, [3]> var_1311 = const()[name = string("op_1311"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1312 = transpose(perm = var_1311, x = x_29_cast_fp16)[name = string("transpose_43")];
+            tensor<fp16, [1, 1536, 1, 1]> input_23 = expand_dims(axes = input_23_axes_0, x = var_1312)[name = string("input_23")];
+            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_25 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = model_model_layers_11_mlp_gate_proj_weight_palettized, x = input_23)[name = string("input_25")];
+            string b_3_pad_type_0 = const()[name = string("b_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_3_strides_0 = const()[name = string("b_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_3_pad_0 = const()[name = string("b_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_3_dilations_0 = const()[name = string("b_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_3_groups_0 = const()[name = string("b_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_3 = conv(dilations = b_3_dilations_0, groups = b_3_groups_0, pad = b_3_pad_0, pad_type = b_3_pad_type_0, strides = b_3_strides_0, weight = model_model_layers_11_mlp_up_proj_weight_palettized, x = input_23)[name = string("b_3")];
+            tensor<fp16, [1, 8960, 1, 1]> c_3 = silu(x = input_25)[name = string("c_3")];
+            tensor<fp16, [1, 8960, 1, 1]> input_27 = mul(x = c_3, y = b_3)[name = string("input_27")];
+            string e_3_pad_type_0 = const()[name = string("e_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_3_strides_0 = const()[name = string("e_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_3_pad_0 = const()[name = string("e_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_3_dilations_0 = const()[name = string("e_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_3_groups_0 = const()[name = string("e_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_3 = conv(dilations = e_3_dilations_0, groups = e_3_groups_0, pad = e_3_pad_0, pad_type = e_3_pad_type_0, strides = e_3_strides_0, weight = model_model_layers_11_mlp_down_proj_weight_palettized, x = input_27)[name = string("e_3")];
+            tensor<int32, [1]> var_1334_axes_0 = const()[name = string("op_1334_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1334 = squeeze(axes = var_1334_axes_0, x = e_3)[name = string("op_1334")];
+            tensor<int32, [3]> var_1335 = const()[name = string("op_1335"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_1336 = transpose(perm = var_1335, x = var_1334)[name = string("transpose_42")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_13_cast_fp16 = add(x = hidden_states_11_cast_fp16, y = var_1336)[name = string("hidden_states_13_cast_fp16")];
+            int32 var_1348 = const()[name = string("op_1348"), val = int32(-1)];
+            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1350_cast_fp16 = mul(x = hidden_states_13_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_1350_cast_fp16")];
+            bool input_29_interleave_0 = const()[name = string("input_29_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_29_cast_fp16 = concat(axis = var_1348, interleave = input_29_interleave_0, values = (hidden_states_13_cast_fp16, var_1350_cast_fp16))[name = string("input_29_cast_fp16")];
+            tensor<int32, [1]> normed_17_axes_0 = const()[name = string("normed_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1345_to_fp16 = const()[name = string("op_1345_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_17_cast_fp16 = layer_norm(axes = normed_17_axes_0, epsilon = var_1345_to_fp16, x = input_29_cast_fp16)[name = string("normed_17_cast_fp16")];
+            tensor<int32, [3]> normed_19_begin_0 = const()[name = string("normed_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_19_end_0 = const()[name = string("normed_19_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_19_end_mask_0 = const()[name = string("normed_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_19_cast_fp16 = slice_by_index(begin = normed_19_begin_0, end = normed_19_end_0, end_mask = normed_19_end_mask_0, x = normed_17_cast_fp16)[name = string("normed_19_cast_fp16")];
+            tensor<fp16, [1536]> const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311911552)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_15_cast_fp16 = mul(x = normed_19_cast_fp16, y = const_47_promoted_to_fp16)[name = string("hidden_states_15_cast_fp16")];
+            tensor<int32, [3]> var_1367 = const()[name = string("op_1367"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_1370_axes_0 = const()[name = string("op_1370_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1368_cast_fp16 = transpose(perm = var_1367, x = hidden_states_15_cast_fp16)[name = string("transpose_41")];
+            tensor<fp16, [1, 1536, 1, 1]> var_1370_cast_fp16 = expand_dims(axes = var_1370_axes_0, x = var_1368_cast_fp16)[name = string("op_1370_cast_fp16")];
+            string var_1386_pad_type_0 = const()[name = string("op_1386_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1386_strides_0 = const()[name = string("op_1386_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1386_pad_0 = const()[name = string("op_1386_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1386_dilations_0 = const()[name = string("op_1386_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1386_groups_0 = const()[name = string("op_1386_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_1386 = conv(bias = model_model_layers_12_self_attn_q_proj_bias, dilations = var_1386_dilations_0, groups = var_1386_groups_0, pad = var_1386_pad_0, pad_type = var_1386_pad_type_0, strides = var_1386_strides_0, weight = model_model_layers_12_self_attn_q_proj_weight_palettized, x = var_1370_cast_fp16)[name = string("op_1386")];
+            tensor<int32, [4]> var_1391 = const()[name = string("op_1391"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_1392 = reshape(shape = var_1391, x = var_1386)[name = string("op_1392")];
+            string var_1408_pad_type_0 = const()[name = string("op_1408_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1408_strides_0 = const()[name = string("op_1408_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1408_pad_0 = const()[name = string("op_1408_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1408_dilations_0 = const()[name = string("op_1408_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1408_groups_0 = const()[name = string("op_1408_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1408 = conv(bias = model_model_layers_12_self_attn_k_proj_bias, dilations = var_1408_dilations_0, groups = var_1408_groups_0, pad = var_1408_pad_0, pad_type = var_1408_pad_type_0, strides = var_1408_strides_0, weight = model_model_layers_12_self_attn_k_proj_weight_palettized, x = var_1370_cast_fp16)[name = string("op_1408")];
+            tensor<int32, [4]> var_1413 = const()[name = string("op_1413"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_1414 = reshape(shape = var_1413, x = var_1408)[name = string("op_1414")];
+            string var_1430_pad_type_0 = const()[name = string("op_1430_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1430_strides_0 = const()[name = string("op_1430_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1430_pad_0 = const()[name = string("op_1430_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1430_dilations_0 = const()[name = string("op_1430_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1430_groups_0 = const()[name = string("op_1430_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1430 = conv(bias = model_model_layers_12_self_attn_v_proj_bias, dilations = var_1430_dilations_0, groups = var_1430_groups_0, pad = var_1430_pad_0, pad_type = var_1430_pad_type_0, strides = var_1430_strides_0, weight = model_model_layers_12_self_attn_v_proj_weight_palettized, x = var_1370_cast_fp16)[name = string("op_1430")];
+            tensor<int32, [4]> var_1435 = const()[name = string("op_1435"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_1436 = reshape(shape = var_1435, x = var_1430)[name = string("op_1436")];
+            tensor<fp16, [1, 12, 1, 128]> var_1442 = mul(x = var_1392, y = cos_1_cast_fp16)[name = string("op_1442")];
+            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_9 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = var_1392)[name = string("x1_9")];
+            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_9 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = var_1392)[name = string("x2_9")];
+            fp16 const_50_promoted = const()[name = string("const_50_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_1463 = mul(x = x2_9, y = const_50_promoted)[name = string("op_1463")];
+            int32 var_1465 = const()[name = string("op_1465"), val = int32(-1)];
+            bool var_1466_interleave_0 = const()[name = string("op_1466_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_1466 = concat(axis = var_1465, interleave = var_1466_interleave_0, values = (var_1463, x1_9))[name = string("op_1466")];
+            tensor<fp16, [1, 12, 1, 128]> var_1467 = mul(x = var_1466, y = sin_1_cast_fp16)[name = string("op_1467")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_5 = add(x = var_1442, y = var_1467)[name = string("query_states_5")];
+            tensor<fp16, [1, 2, 1, 128]> var_1470 = mul(x = var_1414, y = cos_1_cast_fp16)[name = string("op_1470")];
+            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_11 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = var_1414)[name = string("x1_11")];
+            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_11 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = var_1414)[name = string("x2_11")];
+            fp16 const_53_promoted = const()[name = string("const_53_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_1491 = mul(x = x2_11, y = const_53_promoted)[name = string("op_1491")];
+            int32 var_1493 = const()[name = string("op_1493"), val = int32(-1)];
+            bool var_1494_interleave_0 = const()[name = string("op_1494_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_1494 = concat(axis = var_1493, interleave = var_1494_interleave_0, values = (var_1491, x1_11))[name = string("op_1494")];
+            tensor<fp16, [1, 2, 1, 128]> var_1495 = mul(x = var_1494, y = sin_1_cast_fp16)[name = string("op_1495")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_9 = add(x = var_1470, y = var_1495)[name = string("key_states_9")];
+            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([12])];
+            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([13])];
+            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
+            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_24, expand_dims_25, current_pos, expand_dims_27))[name = string("concat_18")];
+            tensor<int32, [1]> concat_19_values1_0 = const()[name = string("concat_19_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_19_values3_0 = const()[name = string("concat_19_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_19_axis_0 = const()[name = string("concat_19_axis_0"), val = int32(0)];
+            bool concat_19_interleave_0 = const()[name = string("concat_19_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_19 = concat(axis = concat_19_axis_0, interleave = concat_19_interleave_0, values = (expand_dims_28, concat_19_values1_0, var_578, concat_19_values3_0))[name = string("concat_19")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_5_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_18, begin_mask = model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0, end = concat_19, end_mask = model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_5_stride_0, update = key_states_9, x = coreml_update_state_21)[name = string("model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_4_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_22 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_4")];
+            tensor<int32, [1]> expand_dims_30 = const()[name = string("expand_dims_30"), val = tensor<int32, [1]>([40])];
+            tensor<int32, [1]> expand_dims_31 = const()[name = string("expand_dims_31"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_33 = const()[name = string("expand_dims_33"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_34 = const()[name = string("expand_dims_34"), val = tensor<int32, [1]>([41])];
+            int32 concat_22_axis_0 = const()[name = string("concat_22_axis_0"), val = int32(0)];
+            bool concat_22_interleave_0 = const()[name = string("concat_22_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_22 = concat(axis = concat_22_axis_0, interleave = concat_22_interleave_0, values = (expand_dims_30, expand_dims_31, current_pos, expand_dims_33))[name = string("concat_22")];
+            tensor<int32, [1]> concat_23_values1_0 = const()[name = string("concat_23_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_23_values3_0 = const()[name = string("concat_23_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_23_axis_0 = const()[name = string("concat_23_axis_0"), val = int32(0)];
+            bool concat_23_interleave_0 = const()[name = string("concat_23_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_23 = concat(axis = concat_23_axis_0, interleave = concat_23_interleave_0, values = (expand_dims_34, concat_23_values1_0, var_578, concat_23_values3_0))[name = string("concat_23")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_6_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_22, begin_mask = model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0, end = concat_23, end_mask = model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_6_stride_0, update = var_1436, x = coreml_update_state_22)[name = string("model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_5_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_23 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_5")];
+            tensor<int32, [4]> var_1550_begin_0 = const()[name = string("op_1550_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> var_1550_end_0 = const()[name = string("op_1550_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
+            tensor<bool, [4]> var_1550_end_mask_0 = const()[name = string("op_1550_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1550_cast_fp16 = slice_by_index(begin = var_1550_begin_0, end = var_1550_end_0, end_mask = var_1550_end_mask_0, x = coreml_update_state_23)[name = string("op_1550_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_5_axes_0 = const()[name = string("K_layer_cache_5_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_5_cast_fp16 = squeeze(axes = K_layer_cache_5_axes_0, x = var_1550_cast_fp16)[name = string("K_layer_cache_5_cast_fp16")];
+            tensor<int32, [4]> var_1557_begin_0 = const()[name = string("op_1557_begin_0"), val = tensor<int32, [4]>([40, 0, 0, 0])];
+            tensor<int32, [4]> var_1557_end_0 = const()[name = string("op_1557_end_0"), val = tensor<int32, [4]>([41, 2, 2048, 128])];
+            tensor<bool, [4]> var_1557_end_mask_0 = const()[name = string("op_1557_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1557_cast_fp16 = slice_by_index(begin = var_1557_begin_0, end = var_1557_end_0, end_mask = var_1557_end_mask_0, x = coreml_update_state_23)[name = string("op_1557_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_5_axes_0 = const()[name = string("V_layer_cache_5_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_5_cast_fp16 = squeeze(axes = V_layer_cache_5_axes_0, x = var_1557_cast_fp16)[name = string("V_layer_cache_5_cast_fp16")];
+            tensor<int32, [1]> x_35_axes_0 = const()[name = string("x_35_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_35_cast_fp16 = expand_dims(axes = x_35_axes_0, x = K_layer_cache_5_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_1594 = const()[name = string("op_1594"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_37_cast_fp16 = tile(reps = var_1594, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_1606 = const()[name = string("op_1606"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_11_cast_fp16 = reshape(shape = var_1606, x = x_37_cast_fp16)[name = string("key_states_11_cast_fp16")];
+            tensor<int32, [1]> x_41_axes_0 = const()[name = string("x_41_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_41_cast_fp16 = expand_dims(axes = x_41_axes_0, x = V_layer_cache_5_cast_fp16)[name = string("x_41_cast_fp16")];
+            tensor<int32, [4]> var_1614 = const()[name = string("op_1614"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_43_cast_fp16 = tile(reps = var_1614, x = x_41_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<int32, [4]> var_1626 = const()[name = string("op_1626"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_11_cast_fp16 = reshape(shape = var_1626, x = x_43_cast_fp16)[name = string("value_states_11_cast_fp16")];
+            bool var_1649_transpose_x_1 = const()[name = string("op_1649_transpose_x_1"), val = bool(false)];
+            bool var_1649_transpose_y_1 = const()[name = string("op_1649_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_1649_cast_fp16 = matmul(transpose_x = var_1649_transpose_x_1, transpose_y = var_1649_transpose_y_1, x = query_states_5, y = key_states_11_cast_fp16)[name = string("op_1649_cast_fp16")];
+            fp16 var_1650_to_fp16 = const()[name = string("op_1650_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_9_cast_fp16 = mul(x = var_1649_cast_fp16, y = var_1650_to_fp16)[name = string("attn_logits_9_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_11_cast_fp16 = add(x = attn_logits_9_cast_fp16, y = causal_mask)[name = string("attn_logits_11_cast_fp16")];
+            int32 var_1677 = const()[name = string("op_1677"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_1679_cast_fp16 = softmax(axis = var_1677, x = attn_logits_11_cast_fp16)[name = string("op_1679_cast_fp16")];
+            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
+            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_1679_cast_fp16, y = value_states_11_cast_fp16)[name = string("attn_output_25_cast_fp16")];
+            tensor<int32, [4]> var_1703_perm_0 = const()[name = string("op_1703_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1707 = const()[name = string("op_1707"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_1703 = transpose(perm = var_1703_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [1, 1, 1536]> attn_output_31 = reshape(shape = var_1707, x = var_1703)[name = string("attn_output_31")];
+            tensor<int32, [3]> var_1712 = const()[name = string("op_1712"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_2_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311914688))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313684224))))[name = string("squeeze_2_palettized")];
+            string var_1728_pad_type_0 = const()[name = string("op_1728_pad_type_0"), val = string("valid")];
+            int32 var_1728_groups_0 = const()[name = string("op_1728_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_1728_strides_0 = const()[name = string("op_1728_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_1728_pad_0 = const()[name = string("op_1728_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_1728_dilations_0 = const()[name = string("op_1728_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_1713 = transpose(perm = var_1712, x = attn_output_31)[name = string("transpose_39")];
+            tensor<fp16, [1, 1536, 1]> var_1728 = conv(dilations = var_1728_dilations_0, groups = var_1728_groups_0, pad = var_1728_pad_0, pad_type = var_1728_pad_type_0, strides = var_1728_strides_0, weight = squeeze_2_palettized, x = var_1713)[name = string("op_1728")];
+            tensor<int32, [3]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_35 = transpose(perm = var_1732, x = var_1728)[name = string("transpose_38")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_17_cast_fp16 = add(x = hidden_states_13_cast_fp16, y = attn_output_35)[name = string("hidden_states_17_cast_fp16")];
+            int32 var_1745 = const()[name = string("op_1745"), val = int32(-1)];
+            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1747_cast_fp16 = mul(x = hidden_states_17_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
+            bool input_35_interleave_0 = const()[name = string("input_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_35_cast_fp16 = concat(axis = var_1745, interleave = input_35_interleave_0, values = (hidden_states_17_cast_fp16, var_1747_cast_fp16))[name = string("input_35_cast_fp16")];
+            tensor<int32, [1]> normed_21_axes_0 = const()[name = string("normed_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1742_to_fp16 = const()[name = string("op_1742_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_21_cast_fp16 = layer_norm(axes = normed_21_axes_0, epsilon = var_1742_to_fp16, x = input_35_cast_fp16)[name = string("normed_21_cast_fp16")];
+            tensor<int32, [3]> normed_23_begin_0 = const()[name = string("normed_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_23_end_0 = const()[name = string("normed_23_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_23_end_mask_0 = const()[name = string("normed_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_23_cast_fp16 = slice_by_index(begin = normed_23_begin_0, end = normed_23_end_0, end_mask = normed_23_end_mask_0, x = normed_21_cast_fp16)[name = string("normed_23_cast_fp16")];
+            tensor<fp16, [1536]> const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313733440)))];
+            tensor<fp16, [1, 1, 1536]> x_45_cast_fp16 = mul(x = normed_23_cast_fp16, y = const_65_promoted_to_fp16)[name = string("x_45_cast_fp16")];
+            tensor<int32, [3]> var_1772 = const()[name = string("op_1772"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_37_axes_0 = const()[name = string("input_37_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1773 = transpose(perm = var_1772, x = x_45_cast_fp16)[name = string("transpose_37")];
+            tensor<fp16, [1, 1536, 1, 1]> input_37 = expand_dims(axes = input_37_axes_0, x = var_1773)[name = string("input_37")];
+            string input_39_pad_type_0 = const()[name = string("input_39_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_39_strides_0 = const()[name = string("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_39_pad_0 = const()[name = string("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_39_dilations_0 = const()[name = string("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_39_groups_0 = const()[name = string("input_39_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_39 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = model_model_layers_12_mlp_gate_proj_weight_palettized, x = input_37)[name = string("input_39")];
+            string b_5_pad_type_0 = const()[name = string("b_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_5_strides_0 = const()[name = string("b_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_5_pad_0 = const()[name = string("b_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_5_dilations_0 = const()[name = string("b_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_5_groups_0 = const()[name = string("b_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_5 = conv(dilations = b_5_dilations_0, groups = b_5_groups_0, pad = b_5_pad_0, pad_type = b_5_pad_type_0, strides = b_5_strides_0, weight = model_model_layers_12_mlp_up_proj_weight_palettized, x = input_37)[name = string("b_5")];
+            tensor<fp16, [1, 8960, 1, 1]> c_5 = silu(x = input_39)[name = string("c_5")];
+            tensor<fp16, [1, 8960, 1, 1]> input_41 = mul(x = c_5, y = b_5)[name = string("input_41")];
+            string e_5_pad_type_0 = const()[name = string("e_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_5_strides_0 = const()[name = string("e_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_5_pad_0 = const()[name = string("e_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_5_dilations_0 = const()[name = string("e_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_5_groups_0 = const()[name = string("e_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_5 = conv(dilations = e_5_dilations_0, groups = e_5_groups_0, pad = e_5_pad_0, pad_type = e_5_pad_type_0, strides = e_5_strides_0, weight = model_model_layers_12_mlp_down_proj_weight_palettized, x = input_41)[name = string("e_5")];
+            tensor<int32, [1]> var_1795_axes_0 = const()[name = string("op_1795_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1795 = squeeze(axes = var_1795_axes_0, x = e_5)[name = string("op_1795")];
+            tensor<int32, [3]> var_1796 = const()[name = string("op_1796"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_1797 = transpose(perm = var_1796, x = var_1795)[name = string("transpose_36")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_19_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = var_1797)[name = string("hidden_states_19_cast_fp16")];
+            int32 var_1809 = const()[name = string("op_1809"), val = int32(-1)];
+            fp16 const_66_promoted_to_fp16 = const()[name = string("const_66_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1811_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_66_promoted_to_fp16)[name = string("op_1811_cast_fp16")];
+            bool input_43_interleave_0 = const()[name = string("input_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_43_cast_fp16 = concat(axis = var_1809, interleave = input_43_interleave_0, values = (hidden_states_19_cast_fp16, var_1811_cast_fp16))[name = string("input_43_cast_fp16")];
+            tensor<int32, [1]> normed_25_axes_0 = const()[name = string("normed_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1806_to_fp16 = const()[name = string("op_1806_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_25_cast_fp16 = layer_norm(axes = normed_25_axes_0, epsilon = var_1806_to_fp16, x = input_43_cast_fp16)[name = string("normed_25_cast_fp16")];
+            tensor<int32, [3]> normed_27_begin_0 = const()[name = string("normed_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_27_end_0 = const()[name = string("normed_27_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_27_end_mask_0 = const()[name = string("normed_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_27_cast_fp16 = slice_by_index(begin = normed_27_begin_0, end = normed_27_end_0, end_mask = normed_27_end_mask_0, x = normed_25_cast_fp16)[name = string("normed_27_cast_fp16")];
+            tensor<fp16, [1536]> const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313736576)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_21_cast_fp16 = mul(x = normed_27_cast_fp16, y = const_69_promoted_to_fp16)[name = string("hidden_states_21_cast_fp16")];
+            tensor<int32, [3]> var_1828 = const()[name = string("op_1828"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_1831_axes_0 = const()[name = string("op_1831_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1829_cast_fp16 = transpose(perm = var_1828, x = hidden_states_21_cast_fp16)[name = string("transpose_35")];
+            tensor<fp16, [1, 1536, 1, 1]> var_1831_cast_fp16 = expand_dims(axes = var_1831_axes_0, x = var_1829_cast_fp16)[name = string("op_1831_cast_fp16")];
+            string var_1847_pad_type_0 = const()[name = string("op_1847_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1847_strides_0 = const()[name = string("op_1847_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1847_pad_0 = const()[name = string("op_1847_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1847_dilations_0 = const()[name = string("op_1847_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1847_groups_0 = const()[name = string("op_1847_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_1847 = conv(bias = model_model_layers_13_self_attn_q_proj_bias, dilations = var_1847_dilations_0, groups = var_1847_groups_0, pad = var_1847_pad_0, pad_type = var_1847_pad_type_0, strides = var_1847_strides_0, weight = model_model_layers_13_self_attn_q_proj_weight_palettized, x = var_1831_cast_fp16)[name = string("op_1847")];
+            tensor<int32, [4]> var_1852 = const()[name = string("op_1852"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_1853 = reshape(shape = var_1852, x = var_1847)[name = string("op_1853")];
+            string var_1869_pad_type_0 = const()[name = string("op_1869_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1869_strides_0 = const()[name = string("op_1869_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1869_pad_0 = const()[name = string("op_1869_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1869_dilations_0 = const()[name = string("op_1869_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1869_groups_0 = const()[name = string("op_1869_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1869 = conv(bias = model_model_layers_13_self_attn_k_proj_bias, dilations = var_1869_dilations_0, groups = var_1869_groups_0, pad = var_1869_pad_0, pad_type = var_1869_pad_type_0, strides = var_1869_strides_0, weight = model_model_layers_13_self_attn_k_proj_weight_palettized, x = var_1831_cast_fp16)[name = string("op_1869")];
+            tensor<int32, [4]> var_1874 = const()[name = string("op_1874"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_1875 = reshape(shape = var_1874, x = var_1869)[name = string("op_1875")];
+            string var_1891_pad_type_0 = const()[name = string("op_1891_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1891_strides_0 = const()[name = string("op_1891_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1891_pad_0 = const()[name = string("op_1891_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1891_dilations_0 = const()[name = string("op_1891_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1891_groups_0 = const()[name = string("op_1891_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1891 = conv(bias = model_model_layers_13_self_attn_v_proj_bias, dilations = var_1891_dilations_0, groups = var_1891_groups_0, pad = var_1891_pad_0, pad_type = var_1891_pad_type_0, strides = var_1891_strides_0, weight = model_model_layers_13_self_attn_v_proj_weight_palettized, x = var_1831_cast_fp16)[name = string("op_1891")];
+            tensor<int32, [4]> var_1896 = const()[name = string("op_1896"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_1897 = reshape(shape = var_1896, x = var_1891)[name = string("op_1897")];
+            tensor<fp16, [1, 12, 1, 128]> var_1903 = mul(x = var_1853, y = cos_1_cast_fp16)[name = string("op_1903")];
+            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_13 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = var_1853)[name = string("x1_13")];
+            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_13 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = var_1853)[name = string("x2_13")];
+            fp16 const_72_promoted = const()[name = string("const_72_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_1924 = mul(x = x2_13, y = const_72_promoted)[name = string("op_1924")];
+            int32 var_1926 = const()[name = string("op_1926"), val = int32(-1)];
+            bool var_1927_interleave_0 = const()[name = string("op_1927_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_1927 = concat(axis = var_1926, interleave = var_1927_interleave_0, values = (var_1924, x1_13))[name = string("op_1927")];
+            tensor<fp16, [1, 12, 1, 128]> var_1928 = mul(x = var_1927, y = sin_1_cast_fp16)[name = string("op_1928")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_7 = add(x = var_1903, y = var_1928)[name = string("query_states_7")];
+            tensor<fp16, [1, 2, 1, 128]> var_1931 = mul(x = var_1875, y = cos_1_cast_fp16)[name = string("op_1931")];
+            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_15 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = var_1875)[name = string("x1_15")];
+            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_15 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = var_1875)[name = string("x2_15")];
+            fp16 const_75_promoted = const()[name = string("const_75_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_1952 = mul(x = x2_15, y = const_75_promoted)[name = string("op_1952")];
+            int32 var_1954 = const()[name = string("op_1954"), val = int32(-1)];
+            bool var_1955_interleave_0 = const()[name = string("op_1955_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_1955 = concat(axis = var_1954, interleave = var_1955_interleave_0, values = (var_1952, x1_15))[name = string("op_1955")];
+            tensor<fp16, [1, 2, 1, 128]> var_1956 = mul(x = var_1955, y = sin_1_cast_fp16)[name = string("op_1956")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_13 = add(x = var_1931, y = var_1956)[name = string("key_states_13")];
+            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([13])];
+            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([14])];
+            int32 concat_26_axis_0 = const()[name = string("concat_26_axis_0"), val = int32(0)];
+            bool concat_26_interleave_0 = const()[name = string("concat_26_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_26 = concat(axis = concat_26_axis_0, interleave = concat_26_interleave_0, values = (expand_dims_36, expand_dims_37, current_pos, expand_dims_39))[name = string("concat_26")];
+            tensor<int32, [1]> concat_27_values1_0 = const()[name = string("concat_27_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_27_values3_0 = const()[name = string("concat_27_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_27_axis_0 = const()[name = string("concat_27_axis_0"), val = int32(0)];
+            bool concat_27_interleave_0 = const()[name = string("concat_27_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_27 = concat(axis = concat_27_axis_0, interleave = concat_27_interleave_0, values = (expand_dims_40, concat_27_values1_0, var_578, concat_27_values3_0))[name = string("concat_27")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_7_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_26, begin_mask = model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0, end = concat_27, end_mask = model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_7_stride_0, update = key_states_13, x = coreml_update_state_23)[name = string("model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_6_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_24 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_6")];
+            tensor<int32, [1]> expand_dims_42 = const()[name = string("expand_dims_42"), val = tensor<int32, [1]>([41])];
+            tensor<int32, [1]> expand_dims_43 = const()[name = string("expand_dims_43"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_45 = const()[name = string("expand_dims_45"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_46 = const()[name = string("expand_dims_46"), val = tensor<int32, [1]>([42])];
+            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
+            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_42, expand_dims_43, current_pos, expand_dims_45))[name = string("concat_30")];
+            tensor<int32, [1]> concat_31_values1_0 = const()[name = string("concat_31_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_31_values3_0 = const()[name = string("concat_31_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_31_axis_0 = const()[name = string("concat_31_axis_0"), val = int32(0)];
+            bool concat_31_interleave_0 = const()[name = string("concat_31_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_31 = concat(axis = concat_31_axis_0, interleave = concat_31_interleave_0, values = (expand_dims_46, concat_31_values1_0, var_578, concat_31_values3_0))[name = string("concat_31")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_8_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_30, begin_mask = model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0, end = concat_31, end_mask = model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_8_stride_0, update = var_1897, x = coreml_update_state_24)[name = string("model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_7_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_25 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_7")];
+            tensor<int32, [4]> var_2011_begin_0 = const()[name = string("op_2011_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> var_2011_end_0 = const()[name = string("op_2011_end_0"), val = tensor<int32, [4]>([14, 2, 2048, 128])];
+            tensor<bool, [4]> var_2011_end_mask_0 = const()[name = string("op_2011_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2011_cast_fp16 = slice_by_index(begin = var_2011_begin_0, end = var_2011_end_0, end_mask = var_2011_end_mask_0, x = coreml_update_state_25)[name = string("op_2011_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_7_axes_0 = const()[name = string("K_layer_cache_7_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_7_cast_fp16 = squeeze(axes = K_layer_cache_7_axes_0, x = var_2011_cast_fp16)[name = string("K_layer_cache_7_cast_fp16")];
+            tensor<int32, [4]> var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor<int32, [4]>([41, 0, 0, 0])];
+            tensor<int32, [4]> var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor<int32, [4]>([42, 2, 2048, 128])];
+            tensor<bool, [4]> var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, x = coreml_update_state_25)[name = string("op_2018_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_7_axes_0 = const()[name = string("V_layer_cache_7_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_7_cast_fp16 = squeeze(axes = V_layer_cache_7_axes_0, x = var_2018_cast_fp16)[name = string("V_layer_cache_7_cast_fp16")];
+            tensor<int32, [1]> x_51_axes_0 = const()[name = string("x_51_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_51_cast_fp16 = expand_dims(axes = x_51_axes_0, x = K_layer_cache_7_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [4]> var_2055 = const()[name = string("op_2055"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_53_cast_fp16 = tile(reps = var_2055, x = x_51_cast_fp16)[name = string("x_53_cast_fp16")];
+            tensor<int32, [4]> var_2067 = const()[name = string("op_2067"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_15_cast_fp16 = reshape(shape = var_2067, x = x_53_cast_fp16)[name = string("key_states_15_cast_fp16")];
+            tensor<int32, [1]> x_57_axes_0 = const()[name = string("x_57_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_57_cast_fp16 = expand_dims(axes = x_57_axes_0, x = V_layer_cache_7_cast_fp16)[name = string("x_57_cast_fp16")];
+            tensor<int32, [4]> var_2075 = const()[name = string("op_2075"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_59_cast_fp16 = tile(reps = var_2075, x = x_57_cast_fp16)[name = string("x_59_cast_fp16")];
+            tensor<int32, [4]> var_2087 = const()[name = string("op_2087"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_15_cast_fp16 = reshape(shape = var_2087, x = x_59_cast_fp16)[name = string("value_states_15_cast_fp16")];
+            bool var_2110_transpose_x_1 = const()[name = string("op_2110_transpose_x_1"), val = bool(false)];
+            bool var_2110_transpose_y_1 = const()[name = string("op_2110_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_1, transpose_y = var_2110_transpose_y_1, x = query_states_7, y = key_states_15_cast_fp16)[name = string("op_2110_cast_fp16")];
+            fp16 var_2111_to_fp16 = const()[name = string("op_2111_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_13_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2111_to_fp16)[name = string("attn_logits_13_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_15_cast_fp16 = add(x = attn_logits_13_cast_fp16, y = causal_mask)[name = string("attn_logits_15_cast_fp16")];
+            int32 var_2138 = const()[name = string("op_2138"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_2140_cast_fp16 = softmax(axis = var_2138, x = attn_logits_15_cast_fp16)[name = string("op_2140_cast_fp16")];
+            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
+            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = var_2140_cast_fp16, y = value_states_15_cast_fp16)[name = string("attn_output_37_cast_fp16")];
+            tensor<int32, [4]> var_2164_perm_0 = const()[name = string("op_2164_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2168 = const()[name = string("op_2168"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_2164 = transpose(perm = var_2164_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_34")];
+            tensor<fp16, [1, 1, 1536]> attn_output_43 = reshape(shape = var_2168, x = var_2164)[name = string("attn_output_43")];
+            tensor<int32, [3]> var_2173 = const()[name = string("op_2173"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_3_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313739712))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315509248))))[name = string("squeeze_3_palettized")];
+            string var_2189_pad_type_0 = const()[name = string("op_2189_pad_type_0"), val = string("valid")];
+            int32 var_2189_groups_0 = const()[name = string("op_2189_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_2189_strides_0 = const()[name = string("op_2189_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_2189_pad_0 = const()[name = string("op_2189_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_2189_dilations_0 = const()[name = string("op_2189_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_2174 = transpose(perm = var_2173, x = attn_output_43)[name = string("transpose_33")];
+            tensor<fp16, [1, 1536, 1]> var_2189 = conv(dilations = var_2189_dilations_0, groups = var_2189_groups_0, pad = var_2189_pad_0, pad_type = var_2189_pad_type_0, strides = var_2189_strides_0, weight = squeeze_3_palettized, x = var_2174)[name = string("op_2189")];
+            tensor<int32, [3]> var_2193 = const()[name = string("op_2193"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_47 = transpose(perm = var_2193, x = var_2189)[name = string("transpose_32")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_23_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = attn_output_47)[name = string("hidden_states_23_cast_fp16")];
+            int32 var_2206 = const()[name = string("op_2206"), val = int32(-1)];
+            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2208_cast_fp16 = mul(x = hidden_states_23_cast_fp16, y = const_84_promoted_to_fp16)[name = string("op_2208_cast_fp16")];
+            bool input_49_interleave_0 = const()[name = string("input_49_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_49_cast_fp16 = concat(axis = var_2206, interleave = input_49_interleave_0, values = (hidden_states_23_cast_fp16, var_2208_cast_fp16))[name = string("input_49_cast_fp16")];
+            tensor<int32, [1]> normed_29_axes_0 = const()[name = string("normed_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2203_to_fp16 = const()[name = string("op_2203_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_29_cast_fp16 = layer_norm(axes = normed_29_axes_0, epsilon = var_2203_to_fp16, x = input_49_cast_fp16)[name = string("normed_29_cast_fp16")];
+            tensor<int32, [3]> normed_31_begin_0 = const()[name = string("normed_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_31_end_0 = const()[name = string("normed_31_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_31_end_mask_0 = const()[name = string("normed_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_31_cast_fp16 = slice_by_index(begin = normed_31_begin_0, end = normed_31_end_0, end_mask = normed_31_end_mask_0, x = normed_29_cast_fp16)[name = string("normed_31_cast_fp16")];
+            tensor<fp16, [1536]> const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315558464)))];
+            tensor<fp16, [1, 1, 1536]> x_61_cast_fp16 = mul(x = normed_31_cast_fp16, y = const_87_promoted_to_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [3]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2234 = transpose(perm = var_2233, x = x_61_cast_fp16)[name = string("transpose_31")];
+            tensor<fp16, [1, 1536, 1, 1]> input_51 = expand_dims(axes = input_51_axes_0, x = var_2234)[name = string("input_51")];
+            string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_53 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = model_model_layers_13_mlp_gate_proj_weight_palettized, x = input_51)[name = string("input_53")];
+            string b_7_pad_type_0 = const()[name = string("b_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_7_strides_0 = const()[name = string("b_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_7_pad_0 = const()[name = string("b_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_7_dilations_0 = const()[name = string("b_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_7_groups_0 = const()[name = string("b_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_7 = conv(dilations = b_7_dilations_0, groups = b_7_groups_0, pad = b_7_pad_0, pad_type = b_7_pad_type_0, strides = b_7_strides_0, weight = model_model_layers_13_mlp_up_proj_weight_palettized, x = input_51)[name = string("b_7")];
+            tensor<fp16, [1, 8960, 1, 1]> c_7 = silu(x = input_53)[name = string("c_7")];
+            tensor<fp16, [1, 8960, 1, 1]> input_55 = mul(x = c_7, y = b_7)[name = string("input_55")];
+            string e_7_pad_type_0 = const()[name = string("e_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_7_strides_0 = const()[name = string("e_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_7_pad_0 = const()[name = string("e_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_7_dilations_0 = const()[name = string("e_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_7_groups_0 = const()[name = string("e_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_7 = conv(dilations = e_7_dilations_0, groups = e_7_groups_0, pad = e_7_pad_0, pad_type = e_7_pad_type_0, strides = e_7_strides_0, weight = model_model_layers_13_mlp_down_proj_weight_palettized, x = input_55)[name = string("e_7")];
+            tensor<int32, [1]> var_2256_axes_0 = const()[name = string("op_2256_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2256 = squeeze(axes = var_2256_axes_0, x = e_7)[name = string("op_2256")];
+            tensor<int32, [3]> var_2257 = const()[name = string("op_2257"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_2258 = transpose(perm = var_2257, x = var_2256)[name = string("transpose_30")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_25_cast_fp16 = add(x = hidden_states_23_cast_fp16, y = var_2258)[name = string("hidden_states_25_cast_fp16")];
+            int32 var_2270 = const()[name = string("op_2270"), val = int32(-1)];
+            fp16 const_88_promoted_to_fp16 = const()[name = string("const_88_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2272_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_88_promoted_to_fp16)[name = string("op_2272_cast_fp16")];
+            bool input_57_interleave_0 = const()[name = string("input_57_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_57_cast_fp16 = concat(axis = var_2270, interleave = input_57_interleave_0, values = (hidden_states_25_cast_fp16, var_2272_cast_fp16))[name = string("input_57_cast_fp16")];
+            tensor<int32, [1]> normed_33_axes_0 = const()[name = string("normed_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2267_to_fp16 = const()[name = string("op_2267_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_33_cast_fp16 = layer_norm(axes = normed_33_axes_0, epsilon = var_2267_to_fp16, x = input_57_cast_fp16)[name = string("normed_33_cast_fp16")];
+            tensor<int32, [3]> normed_35_begin_0 = const()[name = string("normed_35_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_35_end_0 = const()[name = string("normed_35_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_35_end_mask_0 = const()[name = string("normed_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_35_cast_fp16 = slice_by_index(begin = normed_35_begin_0, end = normed_35_end_0, end_mask = normed_35_end_mask_0, x = normed_33_cast_fp16)[name = string("normed_35_cast_fp16")];
+            tensor<fp16, [1536]> const_91_promoted_to_fp16 = const()[name = string("const_91_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315561600)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_27_cast_fp16 = mul(x = normed_35_cast_fp16, y = const_91_promoted_to_fp16)[name = string("hidden_states_27_cast_fp16")];
+            tensor<int32, [3]> var_2289 = const()[name = string("op_2289"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_2292_axes_0 = const()[name = string("op_2292_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2290_cast_fp16 = transpose(perm = var_2289, x = hidden_states_27_cast_fp16)[name = string("transpose_29")];
+            tensor<fp16, [1, 1536, 1, 1]> var_2292_cast_fp16 = expand_dims(axes = var_2292_axes_0, x = var_2290_cast_fp16)[name = string("op_2292_cast_fp16")];
+            string var_2308_pad_type_0 = const()[name = string("op_2308_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2308_strides_0 = const()[name = string("op_2308_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2308_pad_0 = const()[name = string("op_2308_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2308_dilations_0 = const()[name = string("op_2308_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2308_groups_0 = const()[name = string("op_2308_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_2308 = conv(bias = model_model_layers_14_self_attn_q_proj_bias, dilations = var_2308_dilations_0, groups = var_2308_groups_0, pad = var_2308_pad_0, pad_type = var_2308_pad_type_0, strides = var_2308_strides_0, weight = model_model_layers_14_self_attn_q_proj_weight_palettized, x = var_2292_cast_fp16)[name = string("op_2308")];
+            tensor<int32, [4]> var_2313 = const()[name = string("op_2313"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_2314 = reshape(shape = var_2313, x = var_2308)[name = string("op_2314")];
+            string var_2330_pad_type_0 = const()[name = string("op_2330_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2330_strides_0 = const()[name = string("op_2330_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2330_pad_0 = const()[name = string("op_2330_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2330_dilations_0 = const()[name = string("op_2330_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2330_groups_0 = const()[name = string("op_2330_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2330 = conv(bias = model_model_layers_14_self_attn_k_proj_bias, dilations = var_2330_dilations_0, groups = var_2330_groups_0, pad = var_2330_pad_0, pad_type = var_2330_pad_type_0, strides = var_2330_strides_0, weight = model_model_layers_14_self_attn_k_proj_weight_palettized, x = var_2292_cast_fp16)[name = string("op_2330")];
+            tensor<int32, [4]> var_2335 = const()[name = string("op_2335"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_2336 = reshape(shape = var_2335, x = var_2330)[name = string("op_2336")];
+            string var_2352_pad_type_0 = const()[name = string("op_2352_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2352_strides_0 = const()[name = string("op_2352_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2352_pad_0 = const()[name = string("op_2352_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2352_dilations_0 = const()[name = string("op_2352_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2352_groups_0 = const()[name = string("op_2352_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2352 = conv(bias = model_model_layers_14_self_attn_v_proj_bias, dilations = var_2352_dilations_0, groups = var_2352_groups_0, pad = var_2352_pad_0, pad_type = var_2352_pad_type_0, strides = var_2352_strides_0, weight = model_model_layers_14_self_attn_v_proj_weight_palettized, x = var_2292_cast_fp16)[name = string("op_2352")];
+            tensor<int32, [4]> var_2357 = const()[name = string("op_2357"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_2358 = reshape(shape = var_2357, x = var_2352)[name = string("op_2358")];
+            tensor<fp16, [1, 12, 1, 128]> var_2364 = mul(x = var_2314, y = cos_1_cast_fp16)[name = string("op_2364")];
+            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_17 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = var_2314)[name = string("x1_17")];
+            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_17 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = var_2314)[name = string("x2_17")];
+            fp16 const_94_promoted = const()[name = string("const_94_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_2385 = mul(x = x2_17, y = const_94_promoted)[name = string("op_2385")];
+            int32 var_2387 = const()[name = string("op_2387"), val = int32(-1)];
+            bool var_2388_interleave_0 = const()[name = string("op_2388_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_2388 = concat(axis = var_2387, interleave = var_2388_interleave_0, values = (var_2385, x1_17))[name = string("op_2388")];
+            tensor<fp16, [1, 12, 1, 128]> var_2389 = mul(x = var_2388, y = sin_1_cast_fp16)[name = string("op_2389")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_9 = add(x = var_2364, y = var_2389)[name = string("query_states_9")];
+            tensor<fp16, [1, 2, 1, 128]> var_2392 = mul(x = var_2336, y = cos_1_cast_fp16)[name = string("op_2392")];
+            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_19 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = var_2336)[name = string("x1_19")];
+            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_19 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = var_2336)[name = string("x2_19")];
+            fp16 const_97_promoted = const()[name = string("const_97_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_2413 = mul(x = x2_19, y = const_97_promoted)[name = string("op_2413")];
+            int32 var_2415 = const()[name = string("op_2415"), val = int32(-1)];
+            bool var_2416_interleave_0 = const()[name = string("op_2416_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_2416 = concat(axis = var_2415, interleave = var_2416_interleave_0, values = (var_2413, x1_19))[name = string("op_2416")];
+            tensor<fp16, [1, 2, 1, 128]> var_2417 = mul(x = var_2416, y = sin_1_cast_fp16)[name = string("op_2417")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_17 = add(x = var_2392, y = var_2417)[name = string("key_states_17")];
+            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([14])];
+            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([15])];
+            int32 concat_34_axis_0 = const()[name = string("concat_34_axis_0"), val = int32(0)];
+            bool concat_34_interleave_0 = const()[name = string("concat_34_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_34 = concat(axis = concat_34_axis_0, interleave = concat_34_interleave_0, values = (expand_dims_48, expand_dims_49, current_pos, expand_dims_51))[name = string("concat_34")];
+            tensor<int32, [1]> concat_35_values1_0 = const()[name = string("concat_35_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_35_values3_0 = const()[name = string("concat_35_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_35_axis_0 = const()[name = string("concat_35_axis_0"), val = int32(0)];
+            bool concat_35_interleave_0 = const()[name = string("concat_35_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_35 = concat(axis = concat_35_axis_0, interleave = concat_35_interleave_0, values = (expand_dims_52, concat_35_values1_0, var_578, concat_35_values3_0))[name = string("concat_35")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_9_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_9_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_9_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_34, begin_mask = model_model_kv_cache_0_internal_tensor_assign_9_begin_mask_0, end = concat_35, end_mask = model_model_kv_cache_0_internal_tensor_assign_9_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_9_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_9_stride_0, update = key_states_17, x = coreml_update_state_25)[name = string("model_model_kv_cache_0_internal_tensor_assign_9_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_9_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_8_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_26 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_8")];
+            tensor<int32, [1]> expand_dims_54 = const()[name = string("expand_dims_54"), val = tensor<int32, [1]>([42])];
+            tensor<int32, [1]> expand_dims_55 = const()[name = string("expand_dims_55"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_57 = const()[name = string("expand_dims_57"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_58 = const()[name = string("expand_dims_58"), val = tensor<int32, [1]>([43])];
+            int32 concat_38_axis_0 = const()[name = string("concat_38_axis_0"), val = int32(0)];
+            bool concat_38_interleave_0 = const()[name = string("concat_38_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_38 = concat(axis = concat_38_axis_0, interleave = concat_38_interleave_0, values = (expand_dims_54, expand_dims_55, current_pos, expand_dims_57))[name = string("concat_38")];
+            tensor<int32, [1]> concat_39_values1_0 = const()[name = string("concat_39_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_39_values3_0 = const()[name = string("concat_39_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_39_axis_0 = const()[name = string("concat_39_axis_0"), val = int32(0)];
+            bool concat_39_interleave_0 = const()[name = string("concat_39_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_39 = concat(axis = concat_39_axis_0, interleave = concat_39_interleave_0, values = (expand_dims_58, concat_39_values1_0, var_578, concat_39_values3_0))[name = string("concat_39")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_10_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_10_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_10_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_38, begin_mask = model_model_kv_cache_0_internal_tensor_assign_10_begin_mask_0, end = concat_39, end_mask = model_model_kv_cache_0_internal_tensor_assign_10_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_10_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_10_stride_0, update = var_2358, x = coreml_update_state_26)[name = string("model_model_kv_cache_0_internal_tensor_assign_10_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_10_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_9_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_27 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_9")];
+            tensor<int32, [4]> var_2472_begin_0 = const()[name = string("op_2472_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> var_2472_end_0 = const()[name = string("op_2472_end_0"), val = tensor<int32, [4]>([15, 2, 2048, 128])];
+            tensor<bool, [4]> var_2472_end_mask_0 = const()[name = string("op_2472_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2472_cast_fp16 = slice_by_index(begin = var_2472_begin_0, end = var_2472_end_0, end_mask = var_2472_end_mask_0, x = coreml_update_state_27)[name = string("op_2472_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_9_axes_0 = const()[name = string("K_layer_cache_9_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_9_cast_fp16 = squeeze(axes = K_layer_cache_9_axes_0, x = var_2472_cast_fp16)[name = string("K_layer_cache_9_cast_fp16")];
+            tensor<int32, [4]> var_2479_begin_0 = const()[name = string("op_2479_begin_0"), val = tensor<int32, [4]>([42, 0, 0, 0])];
+            tensor<int32, [4]> var_2479_end_0 = const()[name = string("op_2479_end_0"), val = tensor<int32, [4]>([43, 2, 2048, 128])];
+            tensor<bool, [4]> var_2479_end_mask_0 = const()[name = string("op_2479_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2479_cast_fp16 = slice_by_index(begin = var_2479_begin_0, end = var_2479_end_0, end_mask = var_2479_end_mask_0, x = coreml_update_state_27)[name = string("op_2479_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_9_axes_0 = const()[name = string("V_layer_cache_9_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_9_cast_fp16 = squeeze(axes = V_layer_cache_9_axes_0, x = var_2479_cast_fp16)[name = string("V_layer_cache_9_cast_fp16")];
+            tensor<int32, [1]> x_67_axes_0 = const()[name = string("x_67_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_67_cast_fp16 = expand_dims(axes = x_67_axes_0, x = K_layer_cache_9_cast_fp16)[name = string("x_67_cast_fp16")];
+            tensor<int32, [4]> var_2516 = const()[name = string("op_2516"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_69_cast_fp16 = tile(reps = var_2516, x = x_67_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [4]> var_2528 = const()[name = string("op_2528"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_19_cast_fp16 = reshape(shape = var_2528, x = x_69_cast_fp16)[name = string("key_states_19_cast_fp16")];
+            tensor<int32, [1]> x_73_axes_0 = const()[name = string("x_73_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_73_cast_fp16 = expand_dims(axes = x_73_axes_0, x = V_layer_cache_9_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [4]> var_2536 = const()[name = string("op_2536"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_75_cast_fp16 = tile(reps = var_2536, x = x_73_cast_fp16)[name = string("x_75_cast_fp16")];
+            tensor<int32, [4]> var_2548 = const()[name = string("op_2548"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_19_cast_fp16 = reshape(shape = var_2548, x = x_75_cast_fp16)[name = string("value_states_19_cast_fp16")];
+            bool var_2571_transpose_x_1 = const()[name = string("op_2571_transpose_x_1"), val = bool(false)];
+            bool var_2571_transpose_y_1 = const()[name = string("op_2571_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_2571_cast_fp16 = matmul(transpose_x = var_2571_transpose_x_1, transpose_y = var_2571_transpose_y_1, x = query_states_9, y = key_states_19_cast_fp16)[name = string("op_2571_cast_fp16")];
+            fp16 var_2572_to_fp16 = const()[name = string("op_2572_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_17_cast_fp16 = mul(x = var_2571_cast_fp16, y = var_2572_to_fp16)[name = string("attn_logits_17_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_19_cast_fp16 = add(x = attn_logits_17_cast_fp16, y = causal_mask)[name = string("attn_logits_19_cast_fp16")];
+            int32 var_2599 = const()[name = string("op_2599"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_2601_cast_fp16 = softmax(axis = var_2599, x = attn_logits_19_cast_fp16)[name = string("op_2601_cast_fp16")];
+            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
+            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = var_2601_cast_fp16, y = value_states_19_cast_fp16)[name = string("attn_output_49_cast_fp16")];
+            tensor<int32, [4]> var_2625_perm_0 = const()[name = string("op_2625_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2629 = const()[name = string("op_2629"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_2625 = transpose(perm = var_2625_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_28")];
+            tensor<fp16, [1, 1, 1536]> attn_output_55 = reshape(shape = var_2629, x = var_2625)[name = string("attn_output_55")];
+            tensor<int32, [3]> var_2634 = const()[name = string("op_2634"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_4_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315564736))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317334272))))[name = string("squeeze_4_palettized")];
+            string var_2650_pad_type_0 = const()[name = string("op_2650_pad_type_0"), val = string("valid")];
+            int32 var_2650_groups_0 = const()[name = string("op_2650_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_2650_strides_0 = const()[name = string("op_2650_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_2650_pad_0 = const()[name = string("op_2650_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_2650_dilations_0 = const()[name = string("op_2650_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_2635 = transpose(perm = var_2634, x = attn_output_55)[name = string("transpose_27")];
+            tensor<fp16, [1, 1536, 1]> var_2650 = conv(dilations = var_2650_dilations_0, groups = var_2650_groups_0, pad = var_2650_pad_0, pad_type = var_2650_pad_type_0, strides = var_2650_strides_0, weight = squeeze_4_palettized, x = var_2635)[name = string("op_2650")];
+            tensor<int32, [3]> var_2654 = const()[name = string("op_2654"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_59 = transpose(perm = var_2654, x = var_2650)[name = string("transpose_26")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = attn_output_59)[name = string("hidden_states_29_cast_fp16")];
+            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
+            fp16 const_106_promoted_to_fp16 = const()[name = string("const_106_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2669_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_106_promoted_to_fp16)[name = string("op_2669_cast_fp16")];
+            bool input_63_interleave_0 = const()[name = string("input_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_63_cast_fp16 = concat(axis = var_2667, interleave = input_63_interleave_0, values = (hidden_states_29_cast_fp16, var_2669_cast_fp16))[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> normed_37_axes_0 = const()[name = string("normed_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2664_to_fp16 = const()[name = string("op_2664_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_37_cast_fp16 = layer_norm(axes = normed_37_axes_0, epsilon = var_2664_to_fp16, x = input_63_cast_fp16)[name = string("normed_37_cast_fp16")];
+            tensor<int32, [3]> normed_39_begin_0 = const()[name = string("normed_39_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_39_end_0 = const()[name = string("normed_39_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_39_end_mask_0 = const()[name = string("normed_39_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_39_cast_fp16 = slice_by_index(begin = normed_39_begin_0, end = normed_39_end_0, end_mask = normed_39_end_mask_0, x = normed_37_cast_fp16)[name = string("normed_39_cast_fp16")];
+            tensor<fp16, [1536]> const_109_promoted_to_fp16 = const()[name = string("const_109_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317383488)))];
+            tensor<fp16, [1, 1, 1536]> x_77_cast_fp16 = mul(x = normed_39_cast_fp16, y = const_109_promoted_to_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> var_2694 = const()[name = string("op_2694"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2695 = transpose(perm = var_2694, x = x_77_cast_fp16)[name = string("transpose_25")];
+            tensor<fp16, [1, 1536, 1, 1]> input_65 = expand_dims(axes = input_65_axes_0, x = var_2695)[name = string("input_65")];
+            string input_67_pad_type_0 = const()[name = string("input_67_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_67_strides_0 = const()[name = string("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_67_dilations_0 = const()[name = string("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_67_groups_0 = const()[name = string("input_67_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_67 = conv(dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = model_model_layers_14_mlp_gate_proj_weight_palettized, x = input_65)[name = string("input_67")];
+            string b_9_pad_type_0 = const()[name = string("b_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_9_strides_0 = const()[name = string("b_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_9_pad_0 = const()[name = string("b_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_9_dilations_0 = const()[name = string("b_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_9_groups_0 = const()[name = string("b_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_9 = conv(dilations = b_9_dilations_0, groups = b_9_groups_0, pad = b_9_pad_0, pad_type = b_9_pad_type_0, strides = b_9_strides_0, weight = model_model_layers_14_mlp_up_proj_weight_palettized, x = input_65)[name = string("b_9")];
+            tensor<fp16, [1, 8960, 1, 1]> c_9 = silu(x = input_67)[name = string("c_9")];
+            tensor<fp16, [1, 8960, 1, 1]> input_69 = mul(x = c_9, y = b_9)[name = string("input_69")];
+            string e_9_pad_type_0 = const()[name = string("e_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_9_strides_0 = const()[name = string("e_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_9_pad_0 = const()[name = string("e_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_9_dilations_0 = const()[name = string("e_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_9_groups_0 = const()[name = string("e_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_9 = conv(dilations = e_9_dilations_0, groups = e_9_groups_0, pad = e_9_pad_0, pad_type = e_9_pad_type_0, strides = e_9_strides_0, weight = model_model_layers_14_mlp_down_proj_weight_palettized, x = input_69)[name = string("e_9")];
+            tensor<int32, [1]> var_2717_axes_0 = const()[name = string("op_2717_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2717 = squeeze(axes = var_2717_axes_0, x = e_9)[name = string("op_2717")];
+            tensor<int32, [3]> var_2718 = const()[name = string("op_2718"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_2719 = transpose(perm = var_2718, x = var_2717)[name = string("transpose_24")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_31_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = var_2719)[name = string("hidden_states_31_cast_fp16")];
+            int32 var_2731 = const()[name = string("op_2731"), val = int32(-1)];
+            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2733_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_2733_cast_fp16")];
+            bool input_71_interleave_0 = const()[name = string("input_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_71_cast_fp16 = concat(axis = var_2731, interleave = input_71_interleave_0, values = (hidden_states_31_cast_fp16, var_2733_cast_fp16))[name = string("input_71_cast_fp16")];
+            tensor<int32, [1]> normed_41_axes_0 = const()[name = string("normed_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2728_to_fp16 = const()[name = string("op_2728_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_41_cast_fp16 = layer_norm(axes = normed_41_axes_0, epsilon = var_2728_to_fp16, x = input_71_cast_fp16)[name = string("normed_41_cast_fp16")];
+            tensor<int32, [3]> normed_43_begin_0 = const()[name = string("normed_43_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_43_end_0 = const()[name = string("normed_43_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_43_end_mask_0 = const()[name = string("normed_43_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_43_cast_fp16 = slice_by_index(begin = normed_43_begin_0, end = normed_43_end_0, end_mask = normed_43_end_mask_0, x = normed_41_cast_fp16)[name = string("normed_43_cast_fp16")];
+            tensor<fp16, [1536]> const_113_promoted_to_fp16 = const()[name = string("const_113_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317386624)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_33_cast_fp16 = mul(x = normed_43_cast_fp16, y = const_113_promoted_to_fp16)[name = string("hidden_states_33_cast_fp16")];
+            tensor<int32, [3]> var_2750 = const()[name = string("op_2750"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_2753_axes_0 = const()[name = string("op_2753_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2751_cast_fp16 = transpose(perm = var_2750, x = hidden_states_33_cast_fp16)[name = string("transpose_23")];
+            tensor<fp16, [1, 1536, 1, 1]> var_2753_cast_fp16 = expand_dims(axes = var_2753_axes_0, x = var_2751_cast_fp16)[name = string("op_2753_cast_fp16")];
+            string var_2769_pad_type_0 = const()[name = string("op_2769_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2769_strides_0 = const()[name = string("op_2769_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2769_pad_0 = const()[name = string("op_2769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2769_dilations_0 = const()[name = string("op_2769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2769_groups_0 = const()[name = string("op_2769_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_2769 = conv(bias = model_model_layers_15_self_attn_q_proj_bias, dilations = var_2769_dilations_0, groups = var_2769_groups_0, pad = var_2769_pad_0, pad_type = var_2769_pad_type_0, strides = var_2769_strides_0, weight = model_model_layers_15_self_attn_q_proj_weight_palettized, x = var_2753_cast_fp16)[name = string("op_2769")];
+            tensor<int32, [4]> var_2774 = const()[name = string("op_2774"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_2775 = reshape(shape = var_2774, x = var_2769)[name = string("op_2775")];
+            string var_2791_pad_type_0 = const()[name = string("op_2791_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2791_strides_0 = const()[name = string("op_2791_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2791_pad_0 = const()[name = string("op_2791_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2791_dilations_0 = const()[name = string("op_2791_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2791_groups_0 = const()[name = string("op_2791_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2791 = conv(bias = model_model_layers_15_self_attn_k_proj_bias, dilations = var_2791_dilations_0, groups = var_2791_groups_0, pad = var_2791_pad_0, pad_type = var_2791_pad_type_0, strides = var_2791_strides_0, weight = model_model_layers_15_self_attn_k_proj_weight_palettized, x = var_2753_cast_fp16)[name = string("op_2791")];
+            tensor<int32, [4]> var_2796 = const()[name = string("op_2796"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_2797 = reshape(shape = var_2796, x = var_2791)[name = string("op_2797")];
+            string var_2813_pad_type_0 = const()[name = string("op_2813_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2813_strides_0 = const()[name = string("op_2813_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2813_pad_0 = const()[name = string("op_2813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2813_dilations_0 = const()[name = string("op_2813_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2813_groups_0 = const()[name = string("op_2813_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2813 = conv(bias = model_model_layers_15_self_attn_v_proj_bias, dilations = var_2813_dilations_0, groups = var_2813_groups_0, pad = var_2813_pad_0, pad_type = var_2813_pad_type_0, strides = var_2813_strides_0, weight = model_model_layers_15_self_attn_v_proj_weight_palettized, x = var_2753_cast_fp16)[name = string("op_2813")];
+            tensor<int32, [4]> var_2818 = const()[name = string("op_2818"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_2819 = reshape(shape = var_2818, x = var_2813)[name = string("op_2819")];
+            tensor<fp16, [1, 12, 1, 128]> var_2825 = mul(x = var_2775, y = cos_1_cast_fp16)[name = string("op_2825")];
+            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_21 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = var_2775)[name = string("x1_21")];
+            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_21 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = var_2775)[name = string("x2_21")];
+            fp16 const_116_promoted = const()[name = string("const_116_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_2846 = mul(x = x2_21, y = const_116_promoted)[name = string("op_2846")];
+            int32 var_2848 = const()[name = string("op_2848"), val = int32(-1)];
+            bool var_2849_interleave_0 = const()[name = string("op_2849_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_2849 = concat(axis = var_2848, interleave = var_2849_interleave_0, values = (var_2846, x1_21))[name = string("op_2849")];
+            tensor<fp16, [1, 12, 1, 128]> var_2850 = mul(x = var_2849, y = sin_1_cast_fp16)[name = string("op_2850")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_11 = add(x = var_2825, y = var_2850)[name = string("query_states_11")];
+            tensor<fp16, [1, 2, 1, 128]> var_2853 = mul(x = var_2797, y = cos_1_cast_fp16)[name = string("op_2853")];
+            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_23 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = var_2797)[name = string("x1_23")];
+            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_23 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = var_2797)[name = string("x2_23")];
+            fp16 const_119_promoted = const()[name = string("const_119_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_2874 = mul(x = x2_23, y = const_119_promoted)[name = string("op_2874")];
+            int32 var_2876 = const()[name = string("op_2876"), val = int32(-1)];
+            bool var_2877_interleave_0 = const()[name = string("op_2877_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_2877 = concat(axis = var_2876, interleave = var_2877_interleave_0, values = (var_2874, x1_23))[name = string("op_2877")];
+            tensor<fp16, [1, 2, 1, 128]> var_2878 = mul(x = var_2877, y = sin_1_cast_fp16)[name = string("op_2878")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_21 = add(x = var_2853, y = var_2878)[name = string("key_states_21")];
+            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([15])];
+            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([16])];
+            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
+            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_60, expand_dims_61, current_pos, expand_dims_63))[name = string("concat_42")];
+            tensor<int32, [1]> concat_43_values1_0 = const()[name = string("concat_43_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_43_values3_0 = const()[name = string("concat_43_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_43_axis_0 = const()[name = string("concat_43_axis_0"), val = int32(0)];
+            bool concat_43_interleave_0 = const()[name = string("concat_43_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_43 = concat(axis = concat_43_axis_0, interleave = concat_43_interleave_0, values = (expand_dims_64, concat_43_values1_0, var_578, concat_43_values3_0))[name = string("concat_43")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_11_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_11_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_11_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_42, begin_mask = model_model_kv_cache_0_internal_tensor_assign_11_begin_mask_0, end = concat_43, end_mask = model_model_kv_cache_0_internal_tensor_assign_11_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_11_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_11_stride_0, update = key_states_21, x = coreml_update_state_27)[name = string("model_model_kv_cache_0_internal_tensor_assign_11_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_11_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_10_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_28 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_10")];
+            tensor<int32, [1]> expand_dims_66 = const()[name = string("expand_dims_66"), val = tensor<int32, [1]>([43])];
+            tensor<int32, [1]> expand_dims_67 = const()[name = string("expand_dims_67"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_69 = const()[name = string("expand_dims_69"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_70 = const()[name = string("expand_dims_70"), val = tensor<int32, [1]>([44])];
+            int32 concat_46_axis_0 = const()[name = string("concat_46_axis_0"), val = int32(0)];
+            bool concat_46_interleave_0 = const()[name = string("concat_46_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_46 = concat(axis = concat_46_axis_0, interleave = concat_46_interleave_0, values = (expand_dims_66, expand_dims_67, current_pos, expand_dims_69))[name = string("concat_46")];
+            tensor<int32, [1]> concat_47_values1_0 = const()[name = string("concat_47_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_47_values3_0 = const()[name = string("concat_47_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_47_axis_0 = const()[name = string("concat_47_axis_0"), val = int32(0)];
+            bool concat_47_interleave_0 = const()[name = string("concat_47_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_47 = concat(axis = concat_47_axis_0, interleave = concat_47_interleave_0, values = (expand_dims_70, concat_47_values1_0, var_578, concat_47_values3_0))[name = string("concat_47")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_12_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_12_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_12_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_46, begin_mask = model_model_kv_cache_0_internal_tensor_assign_12_begin_mask_0, end = concat_47, end_mask = model_model_kv_cache_0_internal_tensor_assign_12_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_12_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_12_stride_0, update = var_2819, x = coreml_update_state_28)[name = string("model_model_kv_cache_0_internal_tensor_assign_12_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_12_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_11_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_29 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_11")];
+            tensor<int32, [4]> var_2933_begin_0 = const()[name = string("op_2933_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> var_2933_end_0 = const()[name = string("op_2933_end_0"), val = tensor<int32, [4]>([16, 2, 2048, 128])];
+            tensor<bool, [4]> var_2933_end_mask_0 = const()[name = string("op_2933_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2933_cast_fp16 = slice_by_index(begin = var_2933_begin_0, end = var_2933_end_0, end_mask = var_2933_end_mask_0, x = coreml_update_state_29)[name = string("op_2933_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_11_axes_0 = const()[name = string("K_layer_cache_11_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_11_cast_fp16 = squeeze(axes = K_layer_cache_11_axes_0, x = var_2933_cast_fp16)[name = string("K_layer_cache_11_cast_fp16")];
+            tensor<int32, [4]> var_2940_begin_0 = const()[name = string("op_2940_begin_0"), val = tensor<int32, [4]>([43, 0, 0, 0])];
+            tensor<int32, [4]> var_2940_end_0 = const()[name = string("op_2940_end_0"), val = tensor<int32, [4]>([44, 2, 2048, 128])];
+            tensor<bool, [4]> var_2940_end_mask_0 = const()[name = string("op_2940_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2940_cast_fp16 = slice_by_index(begin = var_2940_begin_0, end = var_2940_end_0, end_mask = var_2940_end_mask_0, x = coreml_update_state_29)[name = string("op_2940_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_11_axes_0 = const()[name = string("V_layer_cache_11_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_11_cast_fp16 = squeeze(axes = V_layer_cache_11_axes_0, x = var_2940_cast_fp16)[name = string("V_layer_cache_11_cast_fp16")];
+            tensor<int32, [1]> x_83_axes_0 = const()[name = string("x_83_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_83_cast_fp16 = expand_dims(axes = x_83_axes_0, x = K_layer_cache_11_cast_fp16)[name = string("x_83_cast_fp16")];
+            tensor<int32, [4]> var_2977 = const()[name = string("op_2977"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_85_cast_fp16 = tile(reps = var_2977, x = x_83_cast_fp16)[name = string("x_85_cast_fp16")];
+            tensor<int32, [4]> var_2989 = const()[name = string("op_2989"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_23_cast_fp16 = reshape(shape = var_2989, x = x_85_cast_fp16)[name = string("key_states_23_cast_fp16")];
+            tensor<int32, [1]> x_89_axes_0 = const()[name = string("x_89_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_89_cast_fp16 = expand_dims(axes = x_89_axes_0, x = V_layer_cache_11_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_2997 = const()[name = string("op_2997"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_91_cast_fp16 = tile(reps = var_2997, x = x_89_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<int32, [4]> var_3009 = const()[name = string("op_3009"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_23_cast_fp16 = reshape(shape = var_3009, x = x_91_cast_fp16)[name = string("value_states_23_cast_fp16")];
+            bool var_3032_transpose_x_1 = const()[name = string("op_3032_transpose_x_1"), val = bool(false)];
+            bool var_3032_transpose_y_1 = const()[name = string("op_3032_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_3032_cast_fp16 = matmul(transpose_x = var_3032_transpose_x_1, transpose_y = var_3032_transpose_y_1, x = query_states_11, y = key_states_23_cast_fp16)[name = string("op_3032_cast_fp16")];
+            fp16 var_3033_to_fp16 = const()[name = string("op_3033_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_21_cast_fp16 = mul(x = var_3032_cast_fp16, y = var_3033_to_fp16)[name = string("attn_logits_21_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_23_cast_fp16 = add(x = attn_logits_21_cast_fp16, y = causal_mask)[name = string("attn_logits_23_cast_fp16")];
+            int32 var_3060 = const()[name = string("op_3060"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_3062_cast_fp16 = softmax(axis = var_3060, x = attn_logits_23_cast_fp16)[name = string("op_3062_cast_fp16")];
+            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
+            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = var_3062_cast_fp16, y = value_states_23_cast_fp16)[name = string("attn_output_61_cast_fp16")];
+            tensor<int32, [4]> var_3086_perm_0 = const()[name = string("op_3086_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3090 = const()[name = string("op_3090"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_3086 = transpose(perm = var_3086_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_22")];
+            tensor<fp16, [1, 1, 1536]> attn_output_67 = reshape(shape = var_3090, x = var_3086)[name = string("attn_output_67")];
+            tensor<int32, [3]> var_3095 = const()[name = string("op_3095"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_5_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317389760))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319159296))))[name = string("squeeze_5_palettized")];
+            string var_3111_pad_type_0 = const()[name = string("op_3111_pad_type_0"), val = string("valid")];
+            int32 var_3111_groups_0 = const()[name = string("op_3111_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_3111_strides_0 = const()[name = string("op_3111_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_3111_pad_0 = const()[name = string("op_3111_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_3111_dilations_0 = const()[name = string("op_3111_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_3096 = transpose(perm = var_3095, x = attn_output_67)[name = string("transpose_21")];
+            tensor<fp16, [1, 1536, 1]> var_3111 = conv(dilations = var_3111_dilations_0, groups = var_3111_groups_0, pad = var_3111_pad_0, pad_type = var_3111_pad_type_0, strides = var_3111_strides_0, weight = squeeze_5_palettized, x = var_3096)[name = string("op_3111")];
+            tensor<int32, [3]> var_3115 = const()[name = string("op_3115"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_71 = transpose(perm = var_3115, x = var_3111)[name = string("transpose_20")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_35_cast_fp16 = add(x = hidden_states_31_cast_fp16, y = attn_output_71)[name = string("hidden_states_35_cast_fp16")];
+            int32 var_3128 = const()[name = string("op_3128"), val = int32(-1)];
+            fp16 const_128_promoted_to_fp16 = const()[name = string("const_128_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3130_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_128_promoted_to_fp16)[name = string("op_3130_cast_fp16")];
+            bool input_77_interleave_0 = const()[name = string("input_77_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_77_cast_fp16 = concat(axis = var_3128, interleave = input_77_interleave_0, values = (hidden_states_35_cast_fp16, var_3130_cast_fp16))[name = string("input_77_cast_fp16")];
+            tensor<int32, [1]> normed_45_axes_0 = const()[name = string("normed_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3125_to_fp16 = const()[name = string("op_3125_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_45_cast_fp16 = layer_norm(axes = normed_45_axes_0, epsilon = var_3125_to_fp16, x = input_77_cast_fp16)[name = string("normed_45_cast_fp16")];
+            tensor<int32, [3]> normed_47_begin_0 = const()[name = string("normed_47_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_47_end_0 = const()[name = string("normed_47_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_47_end_mask_0 = const()[name = string("normed_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_47_cast_fp16 = slice_by_index(begin = normed_47_begin_0, end = normed_47_end_0, end_mask = normed_47_end_mask_0, x = normed_45_cast_fp16)[name = string("normed_47_cast_fp16")];
+            tensor<fp16, [1536]> const_131_promoted_to_fp16 = const()[name = string("const_131_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319208512)))];
+            tensor<fp16, [1, 1, 1536]> x_93_cast_fp16 = mul(x = normed_47_cast_fp16, y = const_131_promoted_to_fp16)[name = string("x_93_cast_fp16")];
+            tensor<int32, [3]> var_3155 = const()[name = string("op_3155"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3156 = transpose(perm = var_3155, x = x_93_cast_fp16)[name = string("transpose_19")];
+            tensor<fp16, [1, 1536, 1, 1]> input_79 = expand_dims(axes = input_79_axes_0, x = var_3156)[name = string("input_79")];
+            string input_81_pad_type_0 = const()[name = string("input_81_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_81_strides_0 = const()[name = string("input_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_81_pad_0 = const()[name = string("input_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_81_dilations_0 = const()[name = string("input_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_81_groups_0 = const()[name = string("input_81_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_81 = conv(dilations = input_81_dilations_0, groups = input_81_groups_0, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = input_81_strides_0, weight = model_model_layers_15_mlp_gate_proj_weight_palettized, x = input_79)[name = string("input_81")];
+            string b_11_pad_type_0 = const()[name = string("b_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_11_strides_0 = const()[name = string("b_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_11_pad_0 = const()[name = string("b_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_11_dilations_0 = const()[name = string("b_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_11_groups_0 = const()[name = string("b_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_11 = conv(dilations = b_11_dilations_0, groups = b_11_groups_0, pad = b_11_pad_0, pad_type = b_11_pad_type_0, strides = b_11_strides_0, weight = model_model_layers_15_mlp_up_proj_weight_palettized, x = input_79)[name = string("b_11")];
+            tensor<fp16, [1, 8960, 1, 1]> c_11 = silu(x = input_81)[name = string("c_11")];
+            tensor<fp16, [1, 8960, 1, 1]> input_83 = mul(x = c_11, y = b_11)[name = string("input_83")];
+            string e_11_pad_type_0 = const()[name = string("e_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_11_strides_0 = const()[name = string("e_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_11_pad_0 = const()[name = string("e_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_11_dilations_0 = const()[name = string("e_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_11_groups_0 = const()[name = string("e_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_11 = conv(dilations = e_11_dilations_0, groups = e_11_groups_0, pad = e_11_pad_0, pad_type = e_11_pad_type_0, strides = e_11_strides_0, weight = model_model_layers_15_mlp_down_proj_weight_palettized, x = input_83)[name = string("e_11")];
+            tensor<int32, [1]> var_3178_axes_0 = const()[name = string("op_3178_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3178 = squeeze(axes = var_3178_axes_0, x = e_11)[name = string("op_3178")];
+            tensor<int32, [3]> var_3179 = const()[name = string("op_3179"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_3180 = transpose(perm = var_3179, x = var_3178)[name = string("transpose_18")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_37_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = var_3180)[name = string("hidden_states_37_cast_fp16")];
+            int32 var_3192 = const()[name = string("op_3192"), val = int32(-1)];
+            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3194_cast_fp16 = mul(x = hidden_states_37_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_3194_cast_fp16")];
+            bool input_85_interleave_0 = const()[name = string("input_85_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_85_cast_fp16 = concat(axis = var_3192, interleave = input_85_interleave_0, values = (hidden_states_37_cast_fp16, var_3194_cast_fp16))[name = string("input_85_cast_fp16")];
+            tensor<int32, [1]> normed_49_axes_0 = const()[name = string("normed_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3189_to_fp16 = const()[name = string("op_3189_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_49_cast_fp16 = layer_norm(axes = normed_49_axes_0, epsilon = var_3189_to_fp16, x = input_85_cast_fp16)[name = string("normed_49_cast_fp16")];
+            tensor<int32, [3]> normed_51_begin_0 = const()[name = string("normed_51_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_51_end_0 = const()[name = string("normed_51_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_51_end_mask_0 = const()[name = string("normed_51_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_51_cast_fp16 = slice_by_index(begin = normed_51_begin_0, end = normed_51_end_0, end_mask = normed_51_end_mask_0, x = normed_49_cast_fp16)[name = string("normed_51_cast_fp16")];
+            tensor<fp16, [1536]> const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319211648)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_39_cast_fp16 = mul(x = normed_51_cast_fp16, y = const_135_promoted_to_fp16)[name = string("hidden_states_39_cast_fp16")];
+            tensor<int32, [3]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_3214_axes_0 = const()[name = string("op_3214_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3212_cast_fp16 = transpose(perm = var_3211, x = hidden_states_39_cast_fp16)[name = string("transpose_17")];
+            tensor<fp16, [1, 1536, 1, 1]> var_3214_cast_fp16 = expand_dims(axes = var_3214_axes_0, x = var_3212_cast_fp16)[name = string("op_3214_cast_fp16")];
+            string var_3230_pad_type_0 = const()[name = string("op_3230_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3230_strides_0 = const()[name = string("op_3230_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3230_pad_0 = const()[name = string("op_3230_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3230_dilations_0 = const()[name = string("op_3230_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3230_groups_0 = const()[name = string("op_3230_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_3230 = conv(bias = model_model_layers_16_self_attn_q_proj_bias, dilations = var_3230_dilations_0, groups = var_3230_groups_0, pad = var_3230_pad_0, pad_type = var_3230_pad_type_0, strides = var_3230_strides_0, weight = model_model_layers_16_self_attn_q_proj_weight_palettized, x = var_3214_cast_fp16)[name = string("op_3230")];
+            tensor<int32, [4]> var_3235 = const()[name = string("op_3235"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_3236 = reshape(shape = var_3235, x = var_3230)[name = string("op_3236")];
+            string var_3252_pad_type_0 = const()[name = string("op_3252_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3252_strides_0 = const()[name = string("op_3252_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3252_pad_0 = const()[name = string("op_3252_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3252_dilations_0 = const()[name = string("op_3252_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3252_groups_0 = const()[name = string("op_3252_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_3252 = conv(bias = model_model_layers_16_self_attn_k_proj_bias, dilations = var_3252_dilations_0, groups = var_3252_groups_0, pad = var_3252_pad_0, pad_type = var_3252_pad_type_0, strides = var_3252_strides_0, weight = model_model_layers_16_self_attn_k_proj_weight_palettized, x = var_3214_cast_fp16)[name = string("op_3252")];
+            tensor<int32, [4]> var_3257 = const()[name = string("op_3257"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_3258 = reshape(shape = var_3257, x = var_3252)[name = string("op_3258")];
+            string var_3274_pad_type_0 = const()[name = string("op_3274_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3274_strides_0 = const()[name = string("op_3274_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3274_pad_0 = const()[name = string("op_3274_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3274_dilations_0 = const()[name = string("op_3274_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3274_groups_0 = const()[name = string("op_3274_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_3274 = conv(bias = model_model_layers_16_self_attn_v_proj_bias, dilations = var_3274_dilations_0, groups = var_3274_groups_0, pad = var_3274_pad_0, pad_type = var_3274_pad_type_0, strides = var_3274_strides_0, weight = model_model_layers_16_self_attn_v_proj_weight_palettized, x = var_3214_cast_fp16)[name = string("op_3274")];
+            tensor<int32, [4]> var_3279 = const()[name = string("op_3279"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_3280 = reshape(shape = var_3279, x = var_3274)[name = string("op_3280")];
+            tensor<fp16, [1, 12, 1, 128]> var_3286 = mul(x = var_3236, y = cos_1_cast_fp16)[name = string("op_3286")];
+            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_25 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = var_3236)[name = string("x1_25")];
+            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_25 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = var_3236)[name = string("x2_25")];
+            fp16 const_138_promoted = const()[name = string("const_138_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_3307 = mul(x = x2_25, y = const_138_promoted)[name = string("op_3307")];
+            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
+            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_3310 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3307, x1_25))[name = string("op_3310")];
+            tensor<fp16, [1, 12, 1, 128]> var_3311 = mul(x = var_3310, y = sin_1_cast_fp16)[name = string("op_3311")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_13 = add(x = var_3286, y = var_3311)[name = string("query_states_13")];
+            tensor<fp16, [1, 2, 1, 128]> var_3314 = mul(x = var_3258, y = cos_1_cast_fp16)[name = string("op_3314")];
+            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_27 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = var_3258)[name = string("x1_27")];
+            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_27 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = var_3258)[name = string("x2_27")];
+            fp16 const_141_promoted = const()[name = string("const_141_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_3335 = mul(x = x2_27, y = const_141_promoted)[name = string("op_3335")];
+            int32 var_3337 = const()[name = string("op_3337"), val = int32(-1)];
+            bool var_3338_interleave_0 = const()[name = string("op_3338_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_3338 = concat(axis = var_3337, interleave = var_3338_interleave_0, values = (var_3335, x1_27))[name = string("op_3338")];
+            tensor<fp16, [1, 2, 1, 128]> var_3339 = mul(x = var_3338, y = sin_1_cast_fp16)[name = string("op_3339")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_25 = add(x = var_3314, y = var_3339)[name = string("key_states_25")];
+            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([16])];
+            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([17])];
+            int32 concat_50_axis_0 = const()[name = string("concat_50_axis_0"), val = int32(0)];
+            bool concat_50_interleave_0 = const()[name = string("concat_50_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_50 = concat(axis = concat_50_axis_0, interleave = concat_50_interleave_0, values = (expand_dims_72, expand_dims_73, current_pos, expand_dims_75))[name = string("concat_50")];
+            tensor<int32, [1]> concat_51_values1_0 = const()[name = string("concat_51_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_51_values3_0 = const()[name = string("concat_51_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_51_axis_0 = const()[name = string("concat_51_axis_0"), val = int32(0)];
+            bool concat_51_interleave_0 = const()[name = string("concat_51_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_51 = concat(axis = concat_51_axis_0, interleave = concat_51_interleave_0, values = (expand_dims_76, concat_51_values1_0, var_578, concat_51_values3_0))[name = string("concat_51")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_13_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_13_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_13_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_50, begin_mask = model_model_kv_cache_0_internal_tensor_assign_13_begin_mask_0, end = concat_51, end_mask = model_model_kv_cache_0_internal_tensor_assign_13_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_13_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_13_stride_0, update = key_states_25, x = coreml_update_state_29)[name = string("model_model_kv_cache_0_internal_tensor_assign_13_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_13_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_12_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_30 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_12")];
+            tensor<int32, [1]> expand_dims_78 = const()[name = string("expand_dims_78"), val = tensor<int32, [1]>([44])];
+            tensor<int32, [1]> expand_dims_79 = const()[name = string("expand_dims_79"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_81 = const()[name = string("expand_dims_81"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_82 = const()[name = string("expand_dims_82"), val = tensor<int32, [1]>([45])];
+            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
+            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_78, expand_dims_79, current_pos, expand_dims_81))[name = string("concat_54")];
+            tensor<int32, [1]> concat_55_values1_0 = const()[name = string("concat_55_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_55_values3_0 = const()[name = string("concat_55_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_55_axis_0 = const()[name = string("concat_55_axis_0"), val = int32(0)];
+            bool concat_55_interleave_0 = const()[name = string("concat_55_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_55 = concat(axis = concat_55_axis_0, interleave = concat_55_interleave_0, values = (expand_dims_82, concat_55_values1_0, var_578, concat_55_values3_0))[name = string("concat_55")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_14_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_14_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_14_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_54, begin_mask = model_model_kv_cache_0_internal_tensor_assign_14_begin_mask_0, end = concat_55, end_mask = model_model_kv_cache_0_internal_tensor_assign_14_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_14_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_14_stride_0, update = var_3280, x = coreml_update_state_30)[name = string("model_model_kv_cache_0_internal_tensor_assign_14_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_14_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_13_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_31 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_13")];
+            tensor<int32, [4]> var_3394_begin_0 = const()[name = string("op_3394_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> var_3394_end_0 = const()[name = string("op_3394_end_0"), val = tensor<int32, [4]>([17, 2, 2048, 128])];
+            tensor<bool, [4]> var_3394_end_mask_0 = const()[name = string("op_3394_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3394_cast_fp16 = slice_by_index(begin = var_3394_begin_0, end = var_3394_end_0, end_mask = var_3394_end_mask_0, x = coreml_update_state_31)[name = string("op_3394_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_13_axes_0 = const()[name = string("K_layer_cache_13_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_13_cast_fp16 = squeeze(axes = K_layer_cache_13_axes_0, x = var_3394_cast_fp16)[name = string("K_layer_cache_13_cast_fp16")];
+            tensor<int32, [4]> var_3401_begin_0 = const()[name = string("op_3401_begin_0"), val = tensor<int32, [4]>([44, 0, 0, 0])];
+            tensor<int32, [4]> var_3401_end_0 = const()[name = string("op_3401_end_0"), val = tensor<int32, [4]>([45, 2, 2048, 128])];
+            tensor<bool, [4]> var_3401_end_mask_0 = const()[name = string("op_3401_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3401_cast_fp16 = slice_by_index(begin = var_3401_begin_0, end = var_3401_end_0, end_mask = var_3401_end_mask_0, x = coreml_update_state_31)[name = string("op_3401_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_13_axes_0 = const()[name = string("V_layer_cache_13_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_13_cast_fp16 = squeeze(axes = V_layer_cache_13_axes_0, x = var_3401_cast_fp16)[name = string("V_layer_cache_13_cast_fp16")];
+            tensor<int32, [1]> x_99_axes_0 = const()[name = string("x_99_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_99_cast_fp16 = expand_dims(axes = x_99_axes_0, x = K_layer_cache_13_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [4]> var_3438 = const()[name = string("op_3438"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_101_cast_fp16 = tile(reps = var_3438, x = x_99_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [4]> var_3450 = const()[name = string("op_3450"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_27_cast_fp16 = reshape(shape = var_3450, x = x_101_cast_fp16)[name = string("key_states_27_cast_fp16")];
+            tensor<int32, [1]> x_105_axes_0 = const()[name = string("x_105_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_105_cast_fp16 = expand_dims(axes = x_105_axes_0, x = V_layer_cache_13_cast_fp16)[name = string("x_105_cast_fp16")];
+            tensor<int32, [4]> var_3458 = const()[name = string("op_3458"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_107_cast_fp16 = tile(reps = var_3458, x = x_105_cast_fp16)[name = string("x_107_cast_fp16")];
+            tensor<int32, [4]> var_3470 = const()[name = string("op_3470"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_27_cast_fp16 = reshape(shape = var_3470, x = x_107_cast_fp16)[name = string("value_states_27_cast_fp16")];
+            bool var_3493_transpose_x_1 = const()[name = string("op_3493_transpose_x_1"), val = bool(false)];
+            bool var_3493_transpose_y_1 = const()[name = string("op_3493_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_3493_cast_fp16 = matmul(transpose_x = var_3493_transpose_x_1, transpose_y = var_3493_transpose_y_1, x = query_states_13, y = key_states_27_cast_fp16)[name = string("op_3493_cast_fp16")];
+            fp16 var_3494_to_fp16 = const()[name = string("op_3494_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_25_cast_fp16 = mul(x = var_3493_cast_fp16, y = var_3494_to_fp16)[name = string("attn_logits_25_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_27_cast_fp16 = add(x = attn_logits_25_cast_fp16, y = causal_mask)[name = string("attn_logits_27_cast_fp16")];
+            int32 var_3521 = const()[name = string("op_3521"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_3523_cast_fp16 = softmax(axis = var_3521, x = attn_logits_27_cast_fp16)[name = string("op_3523_cast_fp16")];
+            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
+            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = var_3523_cast_fp16, y = value_states_27_cast_fp16)[name = string("attn_output_73_cast_fp16")];
+            tensor<int32, [4]> var_3547_perm_0 = const()[name = string("op_3547_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3551 = const()[name = string("op_3551"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_3547 = transpose(perm = var_3547_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_16")];
+            tensor<fp16, [1, 1, 1536]> attn_output_79 = reshape(shape = var_3551, x = var_3547)[name = string("attn_output_79")];
+            tensor<int32, [3]> var_3556 = const()[name = string("op_3556"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_6_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319214784))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320984320))))[name = string("squeeze_6_palettized")];
+            string var_3572_pad_type_0 = const()[name = string("op_3572_pad_type_0"), val = string("valid")];
+            int32 var_3572_groups_0 = const()[name = string("op_3572_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_3572_strides_0 = const()[name = string("op_3572_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_3572_pad_0 = const()[name = string("op_3572_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_3572_dilations_0 = const()[name = string("op_3572_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_3557 = transpose(perm = var_3556, x = attn_output_79)[name = string("transpose_15")];
+            tensor<fp16, [1, 1536, 1]> var_3572 = conv(dilations = var_3572_dilations_0, groups = var_3572_groups_0, pad = var_3572_pad_0, pad_type = var_3572_pad_type_0, strides = var_3572_strides_0, weight = squeeze_6_palettized, x = var_3557)[name = string("op_3572")];
+            tensor<int32, [3]> var_3576 = const()[name = string("op_3576"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_83 = transpose(perm = var_3576, x = var_3572)[name = string("transpose_14")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_41_cast_fp16 = add(x = hidden_states_37_cast_fp16, y = attn_output_83)[name = string("hidden_states_41_cast_fp16")];
+            int32 var_3589 = const()[name = string("op_3589"), val = int32(-1)];
+            fp16 const_150_promoted_to_fp16 = const()[name = string("const_150_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3591_cast_fp16 = mul(x = hidden_states_41_cast_fp16, y = const_150_promoted_to_fp16)[name = string("op_3591_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_91_cast_fp16 = concat(axis = var_3589, interleave = input_91_interleave_0, values = (hidden_states_41_cast_fp16, var_3591_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [1]> normed_53_axes_0 = const()[name = string("normed_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_53_cast_fp16 = layer_norm(axes = normed_53_axes_0, epsilon = var_3586_to_fp16, x = input_91_cast_fp16)[name = string("normed_53_cast_fp16")];
+            tensor<int32, [3]> normed_55_begin_0 = const()[name = string("normed_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_55_end_0 = const()[name = string("normed_55_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_55_end_mask_0 = const()[name = string("normed_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_55_cast_fp16 = slice_by_index(begin = normed_55_begin_0, end = normed_55_end_0, end_mask = normed_55_end_mask_0, x = normed_53_cast_fp16)[name = string("normed_55_cast_fp16")];
+            tensor<fp16, [1536]> const_153_promoted_to_fp16 = const()[name = string("const_153_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321033536)))];
+            tensor<fp16, [1, 1, 1536]> x_109_cast_fp16 = mul(x = normed_55_cast_fp16, y = const_153_promoted_to_fp16)[name = string("x_109_cast_fp16")];
+            tensor<int32, [3]> var_3616 = const()[name = string("op_3616"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_93_axes_0 = const()[name = string("input_93_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3617 = transpose(perm = var_3616, x = x_109_cast_fp16)[name = string("transpose_13")];
+            tensor<fp16, [1, 1536, 1, 1]> input_93 = expand_dims(axes = input_93_axes_0, x = var_3617)[name = string("input_93")];
+            string input_95_pad_type_0 = const()[name = string("input_95_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_95_strides_0 = const()[name = string("input_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_95_pad_0 = const()[name = string("input_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_95_dilations_0 = const()[name = string("input_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_95_groups_0 = const()[name = string("input_95_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_95 = conv(dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = model_model_layers_16_mlp_gate_proj_weight_palettized, x = input_93)[name = string("input_95")];
+            string b_13_pad_type_0 = const()[name = string("b_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_13_strides_0 = const()[name = string("b_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_13_pad_0 = const()[name = string("b_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_13_dilations_0 = const()[name = string("b_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_13_groups_0 = const()[name = string("b_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_13 = conv(dilations = b_13_dilations_0, groups = b_13_groups_0, pad = b_13_pad_0, pad_type = b_13_pad_type_0, strides = b_13_strides_0, weight = model_model_layers_16_mlp_up_proj_weight_palettized, x = input_93)[name = string("b_13")];
+            tensor<fp16, [1, 8960, 1, 1]> c_13 = silu(x = input_95)[name = string("c_13")];
+            tensor<fp16, [1, 8960, 1, 1]> input_97 = mul(x = c_13, y = b_13)[name = string("input_97")];
+            string e_13_pad_type_0 = const()[name = string("e_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_13_strides_0 = const()[name = string("e_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_13_pad_0 = const()[name = string("e_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_13_dilations_0 = const()[name = string("e_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_13_groups_0 = const()[name = string("e_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_13 = conv(dilations = e_13_dilations_0, groups = e_13_groups_0, pad = e_13_pad_0, pad_type = e_13_pad_type_0, strides = e_13_strides_0, weight = model_model_layers_16_mlp_down_proj_weight_palettized, x = input_97)[name = string("e_13")];
+            tensor<int32, [1]> var_3639_axes_0 = const()[name = string("op_3639_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3639 = squeeze(axes = var_3639_axes_0, x = e_13)[name = string("op_3639")];
+            tensor<int32, [3]> var_3640 = const()[name = string("op_3640"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_3641 = transpose(perm = var_3640, x = var_3639)[name = string("transpose_12")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_43_cast_fp16 = add(x = hidden_states_41_cast_fp16, y = var_3641)[name = string("hidden_states_43_cast_fp16")];
+            int32 var_3653 = const()[name = string("op_3653"), val = int32(-1)];
+            fp16 const_154_promoted_to_fp16 = const()[name = string("const_154_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3655_cast_fp16 = mul(x = hidden_states_43_cast_fp16, y = const_154_promoted_to_fp16)[name = string("op_3655_cast_fp16")];
+            bool input_99_interleave_0 = const()[name = string("input_99_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_99_cast_fp16 = concat(axis = var_3653, interleave = input_99_interleave_0, values = (hidden_states_43_cast_fp16, var_3655_cast_fp16))[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> normed_57_axes_0 = const()[name = string("normed_57_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_57_cast_fp16 = layer_norm(axes = normed_57_axes_0, epsilon = var_3650_to_fp16, x = input_99_cast_fp16)[name = string("normed_57_cast_fp16")];
+            tensor<int32, [3]> normed_59_begin_0 = const()[name = string("normed_59_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_59_end_0 = const()[name = string("normed_59_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_59_end_mask_0 = const()[name = string("normed_59_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_59_cast_fp16 = slice_by_index(begin = normed_59_begin_0, end = normed_59_end_0, end_mask = normed_59_end_mask_0, x = normed_57_cast_fp16)[name = string("normed_59_cast_fp16")];
+            tensor<fp16, [1536]> const_157_promoted_to_fp16 = const()[name = string("const_157_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321036672)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_45_cast_fp16 = mul(x = normed_59_cast_fp16, y = const_157_promoted_to_fp16)[name = string("hidden_states_45_cast_fp16")];
+            tensor<int32, [3]> var_3672 = const()[name = string("op_3672"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_3675_axes_0 = const()[name = string("op_3675_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3673_cast_fp16 = transpose(perm = var_3672, x = hidden_states_45_cast_fp16)[name = string("transpose_11")];
+            tensor<fp16, [1, 1536, 1, 1]> var_3675_cast_fp16 = expand_dims(axes = var_3675_axes_0, x = var_3673_cast_fp16)[name = string("op_3675_cast_fp16")];
+            string var_3691_pad_type_0 = const()[name = string("op_3691_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3691_strides_0 = const()[name = string("op_3691_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3691_pad_0 = const()[name = string("op_3691_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3691_dilations_0 = const()[name = string("op_3691_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3691_groups_0 = const()[name = string("op_3691_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_3691 = conv(bias = model_model_layers_17_self_attn_q_proj_bias, dilations = var_3691_dilations_0, groups = var_3691_groups_0, pad = var_3691_pad_0, pad_type = var_3691_pad_type_0, strides = var_3691_strides_0, weight = model_model_layers_17_self_attn_q_proj_weight_palettized, x = var_3675_cast_fp16)[name = string("op_3691")];
+            tensor<int32, [4]> var_3696 = const()[name = string("op_3696"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_3697 = reshape(shape = var_3696, x = var_3691)[name = string("op_3697")];
+            string var_3713_pad_type_0 = const()[name = string("op_3713_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3713_strides_0 = const()[name = string("op_3713_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3713_pad_0 = const()[name = string("op_3713_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3713_dilations_0 = const()[name = string("op_3713_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3713_groups_0 = const()[name = string("op_3713_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_3713 = conv(bias = model_model_layers_17_self_attn_k_proj_bias, dilations = var_3713_dilations_0, groups = var_3713_groups_0, pad = var_3713_pad_0, pad_type = var_3713_pad_type_0, strides = var_3713_strides_0, weight = model_model_layers_17_self_attn_k_proj_weight_palettized, x = var_3675_cast_fp16)[name = string("op_3713")];
+            tensor<int32, [4]> var_3718 = const()[name = string("op_3718"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_3719 = reshape(shape = var_3718, x = var_3713)[name = string("op_3719")];
+            string var_3735_pad_type_0 = const()[name = string("op_3735_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3735_strides_0 = const()[name = string("op_3735_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3735_pad_0 = const()[name = string("op_3735_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3735_dilations_0 = const()[name = string("op_3735_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3735_groups_0 = const()[name = string("op_3735_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_3735 = conv(bias = model_model_layers_17_self_attn_v_proj_bias, dilations = var_3735_dilations_0, groups = var_3735_groups_0, pad = var_3735_pad_0, pad_type = var_3735_pad_type_0, strides = var_3735_strides_0, weight = model_model_layers_17_self_attn_v_proj_weight_palettized, x = var_3675_cast_fp16)[name = string("op_3735")];
+            tensor<int32, [4]> var_3740 = const()[name = string("op_3740"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_3741 = reshape(shape = var_3740, x = var_3735)[name = string("op_3741")];
+            tensor<fp16, [1, 12, 1, 128]> var_3747 = mul(x = var_3697, y = cos_1_cast_fp16)[name = string("op_3747")];
+            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_29 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = var_3697)[name = string("x1_29")];
+            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_29 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = var_3697)[name = string("x2_29")];
+            fp16 const_160_promoted = const()[name = string("const_160_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_3768 = mul(x = x2_29, y = const_160_promoted)[name = string("op_3768")];
+            int32 var_3770 = const()[name = string("op_3770"), val = int32(-1)];
+            bool var_3771_interleave_0 = const()[name = string("op_3771_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_3771 = concat(axis = var_3770, interleave = var_3771_interleave_0, values = (var_3768, x1_29))[name = string("op_3771")];
+            tensor<fp16, [1, 12, 1, 128]> var_3772 = mul(x = var_3771, y = sin_1_cast_fp16)[name = string("op_3772")];
+            tensor<fp16, [1, 12, 1, 128]> query_states_15 = add(x = var_3747, y = var_3772)[name = string("query_states_15")];
+            tensor<fp16, [1, 2, 1, 128]> var_3775 = mul(x = var_3719, y = cos_1_cast_fp16)[name = string("op_3775")];
+            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1_31 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = var_3719)[name = string("x1_31")];
+            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2_31 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = var_3719)[name = string("x2_31")];
+            fp16 const_163_promoted = const()[name = string("const_163_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_3796 = mul(x = x2_31, y = const_163_promoted)[name = string("op_3796")];
+            int32 var_3798 = const()[name = string("op_3798"), val = int32(-1)];
+            bool var_3799_interleave_0 = const()[name = string("op_3799_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_3799 = concat(axis = var_3798, interleave = var_3799_interleave_0, values = (var_3796, x1_31))[name = string("op_3799")];
+            tensor<fp16, [1, 2, 1, 128]> var_3800 = mul(x = var_3799, y = sin_1_cast_fp16)[name = string("op_3800")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_29 = add(x = var_3775, y = var_3800)[name = string("key_states_29")];
+            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([17])];
+            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([18])];
+            int32 concat_58_axis_0 = const()[name = string("concat_58_axis_0"), val = int32(0)];
+            bool concat_58_interleave_0 = const()[name = string("concat_58_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_58 = concat(axis = concat_58_axis_0, interleave = concat_58_interleave_0, values = (expand_dims_84, expand_dims_85, current_pos, expand_dims_87))[name = string("concat_58")];
+            tensor<int32, [1]> concat_59_values1_0 = const()[name = string("concat_59_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_59_values3_0 = const()[name = string("concat_59_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_59_axis_0 = const()[name = string("concat_59_axis_0"), val = int32(0)];
+            bool concat_59_interleave_0 = const()[name = string("concat_59_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_59 = concat(axis = concat_59_axis_0, interleave = concat_59_interleave_0, values = (expand_dims_88, concat_59_values1_0, var_578, concat_59_values3_0))[name = string("concat_59")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_15_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_15_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_15_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_58, begin_mask = model_model_kv_cache_0_internal_tensor_assign_15_begin_mask_0, end = concat_59, end_mask = model_model_kv_cache_0_internal_tensor_assign_15_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_15_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_15_stride_0, update = key_states_29, x = coreml_update_state_31)[name = string("model_model_kv_cache_0_internal_tensor_assign_15_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_15_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_14_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_32 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_14")];
+            tensor<int32, [1]> expand_dims_90 = const()[name = string("expand_dims_90"), val = tensor<int32, [1]>([45])];
+            tensor<int32, [1]> expand_dims_91 = const()[name = string("expand_dims_91"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_93 = const()[name = string("expand_dims_93"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_94 = const()[name = string("expand_dims_94"), val = tensor<int32, [1]>([46])];
+            int32 concat_62_axis_0 = const()[name = string("concat_62_axis_0"), val = int32(0)];
+            bool concat_62_interleave_0 = const()[name = string("concat_62_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_62 = concat(axis = concat_62_axis_0, interleave = concat_62_interleave_0, values = (expand_dims_90, expand_dims_91, current_pos, expand_dims_93))[name = string("concat_62")];
+            tensor<int32, [1]> concat_63_values1_0 = const()[name = string("concat_63_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_63_values3_0 = const()[name = string("concat_63_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_63_axis_0 = const()[name = string("concat_63_axis_0"), val = int32(0)];
+            bool concat_63_interleave_0 = const()[name = string("concat_63_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_63 = concat(axis = concat_63_axis_0, interleave = concat_63_interleave_0, values = (expand_dims_94, concat_63_values1_0, var_578, concat_63_values3_0))[name = string("concat_63")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_16_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_16_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_16_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_62, begin_mask = model_model_kv_cache_0_internal_tensor_assign_16_begin_mask_0, end = concat_63, end_mask = model_model_kv_cache_0_internal_tensor_assign_16_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_16_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_16_stride_0, update = var_3741, x = coreml_update_state_32)[name = string("model_model_kv_cache_0_internal_tensor_assign_16_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_16_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_15_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_33 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_15")];
+            tensor<int32, [4]> var_3855_begin_0 = const()[name = string("op_3855_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> var_3855_end_0 = const()[name = string("op_3855_end_0"), val = tensor<int32, [4]>([18, 2, 2048, 128])];
+            tensor<bool, [4]> var_3855_end_mask_0 = const()[name = string("op_3855_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3855_cast_fp16 = slice_by_index(begin = var_3855_begin_0, end = var_3855_end_0, end_mask = var_3855_end_mask_0, x = coreml_update_state_33)[name = string("op_3855_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_15_axes_0 = const()[name = string("K_layer_cache_15_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_15_cast_fp16 = squeeze(axes = K_layer_cache_15_axes_0, x = var_3855_cast_fp16)[name = string("K_layer_cache_15_cast_fp16")];
+            tensor<int32, [4]> var_3862_begin_0 = const()[name = string("op_3862_begin_0"), val = tensor<int32, [4]>([45, 0, 0, 0])];
+            tensor<int32, [4]> var_3862_end_0 = const()[name = string("op_3862_end_0"), val = tensor<int32, [4]>([46, 2, 2048, 128])];
+            tensor<bool, [4]> var_3862_end_mask_0 = const()[name = string("op_3862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3862_cast_fp16 = slice_by_index(begin = var_3862_begin_0, end = var_3862_end_0, end_mask = var_3862_end_mask_0, x = coreml_update_state_33)[name = string("op_3862_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_15_axes_0 = const()[name = string("V_layer_cache_15_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_15_cast_fp16 = squeeze(axes = V_layer_cache_15_axes_0, x = var_3862_cast_fp16)[name = string("V_layer_cache_15_cast_fp16")];
+            tensor<int32, [1]> x_115_axes_0 = const()[name = string("x_115_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_115_cast_fp16 = expand_dims(axes = x_115_axes_0, x = K_layer_cache_15_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_3899 = const()[name = string("op_3899"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_117_cast_fp16 = tile(reps = var_3899, x = x_115_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_3911 = const()[name = string("op_3911"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_31_cast_fp16 = reshape(shape = var_3911, x = x_117_cast_fp16)[name = string("key_states_31_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_121_cast_fp16 = expand_dims(axes = x_121_axes_0, x = V_layer_cache_15_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [4]> var_3919 = const()[name = string("op_3919"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_123_cast_fp16 = tile(reps = var_3919, x = x_121_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<int32, [4]> var_3931 = const()[name = string("op_3931"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_31_cast_fp16 = reshape(shape = var_3931, x = x_123_cast_fp16)[name = string("value_states_31_cast_fp16")];
+            bool var_3954_transpose_x_1 = const()[name = string("op_3954_transpose_x_1"), val = bool(false)];
+            bool var_3954_transpose_y_1 = const()[name = string("op_3954_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_3954_cast_fp16 = matmul(transpose_x = var_3954_transpose_x_1, transpose_y = var_3954_transpose_y_1, x = query_states_15, y = key_states_31_cast_fp16)[name = string("op_3954_cast_fp16")];
+            fp16 var_3955_to_fp16 = const()[name = string("op_3955_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_29_cast_fp16 = mul(x = var_3954_cast_fp16, y = var_3955_to_fp16)[name = string("attn_logits_29_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_31_cast_fp16 = add(x = attn_logits_29_cast_fp16, y = causal_mask)[name = string("attn_logits_31_cast_fp16")];
+            int32 var_3982 = const()[name = string("op_3982"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_3984_cast_fp16 = softmax(axis = var_3982, x = attn_logits_31_cast_fp16)[name = string("op_3984_cast_fp16")];
+            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
+            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = var_3984_cast_fp16, y = value_states_31_cast_fp16)[name = string("attn_output_85_cast_fp16")];
+            tensor<int32, [4]> var_4008_perm_0 = const()[name = string("op_4008_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4012 = const()[name = string("op_4012"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_4008 = transpose(perm = var_4008_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_10")];
+            tensor<fp16, [1, 1, 1536]> attn_output_91 = reshape(shape = var_4012, x = var_4008)[name = string("attn_output_91")];
+            tensor<int32, [3]> var_4017 = const()[name = string("op_4017"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_7_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321039808))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322809344))))[name = string("squeeze_7_palettized")];
+            string var_4033_pad_type_0 = const()[name = string("op_4033_pad_type_0"), val = string("valid")];
+            int32 var_4033_groups_0 = const()[name = string("op_4033_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_4033_strides_0 = const()[name = string("op_4033_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_4033_pad_0 = const()[name = string("op_4033_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_4033_dilations_0 = const()[name = string("op_4033_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_4018 = transpose(perm = var_4017, x = attn_output_91)[name = string("transpose_9")];
+            tensor<fp16, [1, 1536, 1]> var_4033 = conv(dilations = var_4033_dilations_0, groups = var_4033_groups_0, pad = var_4033_pad_0, pad_type = var_4033_pad_type_0, strides = var_4033_strides_0, weight = squeeze_7_palettized, x = var_4018)[name = string("op_4033")];
+            tensor<int32, [3]> var_4037 = const()[name = string("op_4037"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output_95 = transpose(perm = var_4037, x = var_4033)[name = string("transpose_8")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_47_cast_fp16 = add(x = hidden_states_43_cast_fp16, y = attn_output_95)[name = string("hidden_states_47_cast_fp16")];
+            int32 var_4050 = const()[name = string("op_4050"), val = int32(-1)];
+            fp16 const_172_promoted_to_fp16 = const()[name = string("const_172_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4052_cast_fp16 = mul(x = hidden_states_47_cast_fp16, y = const_172_promoted_to_fp16)[name = string("op_4052_cast_fp16")];
+            bool input_105_interleave_0 = const()[name = string("input_105_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_105_cast_fp16 = concat(axis = var_4050, interleave = input_105_interleave_0, values = (hidden_states_47_cast_fp16, var_4052_cast_fp16))[name = string("input_105_cast_fp16")];
+            tensor<int32, [1]> normed_61_axes_0 = const()[name = string("normed_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4047_to_fp16 = const()[name = string("op_4047_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_61_cast_fp16 = layer_norm(axes = normed_61_axes_0, epsilon = var_4047_to_fp16, x = input_105_cast_fp16)[name = string("normed_61_cast_fp16")];
+            tensor<int32, [3]> normed_63_begin_0 = const()[name = string("normed_63_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_63_end_0 = const()[name = string("normed_63_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_63_end_mask_0 = const()[name = string("normed_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_63_cast_fp16 = slice_by_index(begin = normed_63_begin_0, end = normed_63_end_0, end_mask = normed_63_end_mask_0, x = normed_61_cast_fp16)[name = string("normed_63_cast_fp16")];
+            tensor<fp16, [1536]> const_175_promoted_to_fp16 = const()[name = string("const_175_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322858560)))];
+            tensor<fp16, [1, 1, 1536]> x_125_cast_fp16 = mul(x = normed_63_cast_fp16, y = const_175_promoted_to_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> var_4077 = const()[name = string("op_4077"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4078 = transpose(perm = var_4077, x = x_125_cast_fp16)[name = string("transpose_7")];
+            tensor<fp16, [1, 1536, 1, 1]> input_107 = expand_dims(axes = input_107_axes_0, x = var_4078)[name = string("input_107")];
+            string input_109_pad_type_0 = const()[name = string("input_109_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_109_strides_0 = const()[name = string("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_109_pad_0 = const()[name = string("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_109_dilations_0 = const()[name = string("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_109_groups_0 = const()[name = string("input_109_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_109 = conv(dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = model_model_layers_17_mlp_gate_proj_weight_palettized, x = input_107)[name = string("input_109")];
+            string b_15_pad_type_0 = const()[name = string("b_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_15_strides_0 = const()[name = string("b_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_15_pad_0 = const()[name = string("b_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_15_dilations_0 = const()[name = string("b_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_15_groups_0 = const()[name = string("b_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b_15 = conv(dilations = b_15_dilations_0, groups = b_15_groups_0, pad = b_15_pad_0, pad_type = b_15_pad_type_0, strides = b_15_strides_0, weight = model_model_layers_17_mlp_up_proj_weight_palettized, x = input_107)[name = string("b_15")];
+            tensor<fp16, [1, 8960, 1, 1]> c_15 = silu(x = input_109)[name = string("c_15")];
+            tensor<fp16, [1, 8960, 1, 1]> input_111 = mul(x = c_15, y = b_15)[name = string("input_111")];
+            string e_15_pad_type_0 = const()[name = string("e_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_15_strides_0 = const()[name = string("e_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_15_pad_0 = const()[name = string("e_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_15_dilations_0 = const()[name = string("e_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_15_groups_0 = const()[name = string("e_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e_15 = conv(dilations = e_15_dilations_0, groups = e_15_groups_0, pad = e_15_pad_0, pad_type = e_15_pad_type_0, strides = e_15_strides_0, weight = model_model_layers_17_mlp_down_proj_weight_palettized, x = input_111)[name = string("e_15")];
+            tensor<int32, [1]> var_4100_axes_0 = const()[name = string("op_4100_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4100 = squeeze(axes = var_4100_axes_0, x = e_15)[name = string("op_4100")];
+            tensor<int32, [3]> var_4101 = const()[name = string("op_4101"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_4102 = transpose(perm = var_4101, x = var_4100)[name = string("transpose_6")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_49_cast_fp16 = add(x = hidden_states_47_cast_fp16, y = var_4102)[name = string("hidden_states_49_cast_fp16")];
+            int32 var_4114 = const()[name = string("op_4114"), val = int32(-1)];
+            fp16 const_176_promoted_to_fp16 = const()[name = string("const_176_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4116_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_176_promoted_to_fp16)[name = string("op_4116_cast_fp16")];
+            bool input_113_interleave_0 = const()[name = string("input_113_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_113_cast_fp16 = concat(axis = var_4114, interleave = input_113_interleave_0, values = (hidden_states_49_cast_fp16, var_4116_cast_fp16))[name = string("input_113_cast_fp16")];
+            tensor<int32, [1]> normed_65_axes_0 = const()[name = string("normed_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4111_to_fp16 = const()[name = string("op_4111_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_65_cast_fp16 = layer_norm(axes = normed_65_axes_0, epsilon = var_4111_to_fp16, x = input_113_cast_fp16)[name = string("normed_65_cast_fp16")];
+            tensor<int32, [3]> normed_67_begin_0 = const()[name = string("normed_67_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_67_end_0 = const()[name = string("normed_67_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_67_end_mask_0 = const()[name = string("normed_67_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_67_cast_fp16 = slice_by_index(begin = normed_67_begin_0, end = normed_67_end_0, end_mask = normed_67_end_mask_0, x = normed_65_cast_fp16)[name = string("normed_67_cast_fp16")];
+            tensor<fp16, [1536]> const_179_promoted_to_fp16 = const()[name = string("const_179_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322861696)))];
+            tensor<fp16, [1, 1, 1536]> hidden_states_51_cast_fp16 = mul(x = normed_67_cast_fp16, y = const_179_promoted_to_fp16)[name = string("hidden_states_51_cast_fp16")];
+            tensor<int32, [3]> var_4133 = const()[name = string("op_4133"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_4136_axes_0 = const()[name = string("op_4136_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4134_cast_fp16 = transpose(perm = var_4133, x = hidden_states_51_cast_fp16)[name = string("transpose_5")];
+            tensor<fp16, [1, 1536, 1, 1]> var_4136_cast_fp16 = expand_dims(axes = var_4136_axes_0, x = var_4134_cast_fp16)[name = string("op_4136_cast_fp16")];
+            string var_4152_pad_type_0 = const()[name = string("op_4152_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4152_strides_0 = const()[name = string("op_4152_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4152_pad_0 = const()[name = string("op_4152_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4152_dilations_0 = const()[name = string("op_4152_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4152_groups_0 = const()[name = string("op_4152_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_4152 = conv(bias = model_model_layers_18_self_attn_q_proj_bias, dilations = var_4152_dilations_0, groups = var_4152_groups_0, pad = var_4152_pad_0, pad_type = var_4152_pad_type_0, strides = var_4152_strides_0, weight = model_model_layers_18_self_attn_q_proj_weight_palettized, x = var_4136_cast_fp16)[name = string("op_4152")];
+            tensor<int32, [4]> var_4157 = const()[name = string("op_4157"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<fp16, [1, 12, 1, 128]> var_4158 = reshape(shape = var_4157, x = var_4152)[name = string("op_4158")];
+            string var_4174_pad_type_0 = const()[name = string("op_4174_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4174_strides_0 = const()[name = string("op_4174_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4174_pad_0 = const()[name = string("op_4174_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4174_dilations_0 = const()[name = string("op_4174_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4174_groups_0 = const()[name = string("op_4174_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_4174 = conv(bias = model_model_layers_18_self_attn_k_proj_bias, dilations = var_4174_dilations_0, groups = var_4174_groups_0, pad = var_4174_pad_0, pad_type = var_4174_pad_type_0, strides = var_4174_strides_0, weight = model_model_layers_18_self_attn_k_proj_weight_palettized, x = var_4136_cast_fp16)[name = string("op_4174")];
+            tensor<int32, [4]> var_4179 = const()[name = string("op_4179"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_4180 = reshape(shape = var_4179, x = var_4174)[name = string("op_4180")];
+            string var_4196_pad_type_0 = const()[name = string("op_4196_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4196_strides_0 = const()[name = string("op_4196_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4196_pad_0 = const()[name = string("op_4196_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4196_dilations_0 = const()[name = string("op_4196_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4196_groups_0 = const()[name = string("op_4196_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_4196 = conv(bias = model_model_layers_18_self_attn_v_proj_bias, dilations = var_4196_dilations_0, groups = var_4196_groups_0, pad = var_4196_pad_0, pad_type = var_4196_pad_type_0, strides = var_4196_strides_0, weight = model_model_layers_18_self_attn_v_proj_weight_palettized, x = var_4136_cast_fp16)[name = string("op_4196")];
+            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<fp16, [1, 2, 1, 128]> var_4202 = reshape(shape = var_4201, x = var_4196)[name = string("op_4202")];
+            tensor<fp16, [1, 12, 1, 128]> var_4208 = mul(x = var_4158, y = cos_1_cast_fp16)[name = string("op_4208")];
+            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
+            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 1, 64]> x1_33 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = var_4158)[name = string("x1_33")];
+            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
+            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 1, 64]> x2_33 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = var_4158)[name = string("x2_33")];
+            fp16 const_182_promoted = const()[name = string("const_182_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 1, 64]> var_4229 = mul(x = x2_33, y = const_182_promoted)[name = string("op_4229")];
+            int32 var_4231 = const()[name = string("op_4231"), val = int32(-1)];
+            bool var_4232_interleave_0 = const()[name = string("op_4232_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> var_4232 = concat(axis = var_4231, interleave = var_4232_interleave_0, values = (var_4229, x1_33))[name = string("op_4232")];
+            tensor<fp16, [1, 12, 1, 128]> var_4233 = mul(x = var_4232, y = sin_1_cast_fp16)[name = string("op_4233")];
+            tensor<fp16, [1, 12, 1, 128]> query_states = add(x = var_4208, y = var_4233)[name = string("query_states")];
+            tensor<fp16, [1, 2, 1, 128]> var_4236 = mul(x = var_4180, y = cos_1_cast_fp16)[name = string("op_4236")];
+            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 1, 64])];
+            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 1, 64]> x1 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = var_4180)[name = string("x1")];
+            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 1, 128])];
+            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 1, 64]> x2 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = var_4180)[name = string("x2")];
+            fp16 const_185_promoted = const()[name = string("const_185_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 1, 64]> var_4257 = mul(x = x2, y = const_185_promoted)[name = string("op_4257")];
+            int32 var_4259 = const()[name = string("op_4259"), val = int32(-1)];
+            bool var_4260_interleave_0 = const()[name = string("op_4260_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 1, 128]> var_4260 = concat(axis = var_4259, interleave = var_4260_interleave_0, values = (var_4257, x1))[name = string("op_4260")];
+            tensor<fp16, [1, 2, 1, 128]> var_4261 = mul(x = var_4260, y = sin_1_cast_fp16)[name = string("op_4261")];
+            tensor<fp16, [1, 2, 1, 128]> key_states_33 = add(x = var_4236, y = var_4261)[name = string("key_states_33")];
+            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([18])];
+            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([19])];
+            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
+            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_96, expand_dims_97, current_pos, expand_dims_99))[name = string("concat_66")];
+            tensor<int32, [1]> concat_67_values1_0 = const()[name = string("concat_67_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_67_values3_0 = const()[name = string("concat_67_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_67_axis_0 = const()[name = string("concat_67_axis_0"), val = int32(0)];
+            bool concat_67_interleave_0 = const()[name = string("concat_67_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_67 = concat(axis = concat_67_axis_0, interleave = concat_67_interleave_0, values = (expand_dims_100, concat_67_values1_0, var_578, concat_67_values3_0))[name = string("concat_67")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_17_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_17_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_17_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_66, begin_mask = model_model_kv_cache_0_internal_tensor_assign_17_begin_mask_0, end = concat_67, end_mask = model_model_kv_cache_0_internal_tensor_assign_17_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_17_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_17_stride_0, update = key_states_33, x = coreml_update_state_33)[name = string("model_model_kv_cache_0_internal_tensor_assign_17_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_17_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_16_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_34 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_16")];
+            tensor<int32, [1]> expand_dims_102 = const()[name = string("expand_dims_102"), val = tensor<int32, [1]>([46])];
+            tensor<int32, [1]> expand_dims_103 = const()[name = string("expand_dims_103"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_105 = const()[name = string("expand_dims_105"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_106 = const()[name = string("expand_dims_106"), val = tensor<int32, [1]>([47])];
+            int32 concat_70_axis_0 = const()[name = string("concat_70_axis_0"), val = int32(0)];
+            bool concat_70_interleave_0 = const()[name = string("concat_70_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_70 = concat(axis = concat_70_axis_0, interleave = concat_70_interleave_0, values = (expand_dims_102, expand_dims_103, current_pos, expand_dims_105))[name = string("concat_70")];
+            tensor<int32, [1]> concat_71_values1_0 = const()[name = string("concat_71_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_71_values3_0 = const()[name = string("concat_71_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_71_axis_0 = const()[name = string("concat_71_axis_0"), val = int32(0)];
+            bool concat_71_interleave_0 = const()[name = string("concat_71_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_71 = concat(axis = concat_71_axis_0, interleave = concat_71_interleave_0, values = (expand_dims_106, concat_71_values1_0, var_578, concat_71_values3_0))[name = string("concat_71")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_18_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_18_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_18_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_70, begin_mask = model_model_kv_cache_0_internal_tensor_assign_18_begin_mask_0, end = concat_71, end_mask = model_model_kv_cache_0_internal_tensor_assign_18_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_18_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_18_stride_0, update = var_4202, x = coreml_update_state_34)[name = string("model_model_kv_cache_0_internal_tensor_assign_18_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_18_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_17_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_35 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_17")];
+            tensor<int32, [4]> var_4316_begin_0 = const()[name = string("op_4316_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> var_4316_end_0 = const()[name = string("op_4316_end_0"), val = tensor<int32, [4]>([19, 2, 2048, 128])];
+            tensor<bool, [4]> var_4316_end_mask_0 = const()[name = string("op_4316_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_4316_cast_fp16 = slice_by_index(begin = var_4316_begin_0, end = var_4316_end_0, end_mask = var_4316_end_mask_0, x = coreml_update_state_35)[name = string("op_4316_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_axes_0 = const()[name = string("K_layer_cache_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_cast_fp16 = squeeze(axes = K_layer_cache_axes_0, x = var_4316_cast_fp16)[name = string("K_layer_cache_cast_fp16")];
+            tensor<int32, [4]> var_4323_begin_0 = const()[name = string("op_4323_begin_0"), val = tensor<int32, [4]>([46, 0, 0, 0])];
+            tensor<int32, [4]> var_4323_end_0 = const()[name = string("op_4323_end_0"), val = tensor<int32, [4]>([47, 2, 2048, 128])];
+            tensor<bool, [4]> var_4323_end_mask_0 = const()[name = string("op_4323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_4323_cast_fp16 = slice_by_index(begin = var_4323_begin_0, end = var_4323_end_0, end_mask = var_4323_end_mask_0, x = coreml_update_state_35)[name = string("op_4323_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_axes_0 = const()[name = string("V_layer_cache_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_cast_fp16 = squeeze(axes = V_layer_cache_axes_0, x = var_4323_cast_fp16)[name = string("V_layer_cache_cast_fp16")];
+            tensor<int32, [1]> x_131_axes_0 = const()[name = string("x_131_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_131_cast_fp16 = expand_dims(axes = x_131_axes_0, x = K_layer_cache_cast_fp16)[name = string("x_131_cast_fp16")];
+            tensor<int32, [4]> var_4360 = const()[name = string("op_4360"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_133_cast_fp16 = tile(reps = var_4360, x = x_131_cast_fp16)[name = string("x_133_cast_fp16")];
+            tensor<int32, [4]> var_4372 = const()[name = string("op_4372"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_cast_fp16 = reshape(shape = var_4372, x = x_133_cast_fp16)[name = string("key_states_cast_fp16")];
+            tensor<int32, [1]> x_137_axes_0 = const()[name = string("x_137_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_137_cast_fp16 = expand_dims(axes = x_137_axes_0, x = V_layer_cache_cast_fp16)[name = string("x_137_cast_fp16")];
+            tensor<int32, [4]> var_4380 = const()[name = string("op_4380"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_139_cast_fp16 = tile(reps = var_4380, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            tensor<int32, [4]> var_4392 = const()[name = string("op_4392"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> value_states_cast_fp16 = reshape(shape = var_4392, x = x_139_cast_fp16)[name = string("value_states_cast_fp16")];
+            bool var_4415_transpose_x_1 = const()[name = string("op_4415_transpose_x_1"), val = bool(false)];
+            bool var_4415_transpose_y_1 = const()[name = string("op_4415_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 1, 2048]> var_4415_cast_fp16 = matmul(transpose_x = var_4415_transpose_x_1, transpose_y = var_4415_transpose_y_1, x = query_states, y = key_states_cast_fp16)[name = string("op_4415_cast_fp16")];
+            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_33_cast_fp16 = mul(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("attn_logits_33_cast_fp16")];
+            tensor<fp16, [1, 12, 1, 2048]> attn_logits_cast_fp16 = add(x = attn_logits_33_cast_fp16, y = causal_mask)[name = string("attn_logits_cast_fp16")];
+            int32 var_4443 = const()[name = string("op_4443"), val = int32(-1)];
+            tensor<fp16, [1, 12, 1, 2048]> var_4445_cast_fp16 = softmax(axis = var_4443, x = attn_logits_cast_fp16)[name = string("op_4445_cast_fp16")];
+            bool attn_output_97_transpose_x_0 = const()[name = string("attn_output_97_transpose_x_0"), val = bool(false)];
+            bool attn_output_97_transpose_y_0 = const()[name = string("attn_output_97_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 1, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = var_4445_cast_fp16, y = value_states_cast_fp16)[name = string("attn_output_97_cast_fp16")];
+            tensor<int32, [4]> var_4469_perm_0 = const()[name = string("op_4469_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4473 = const()[name = string("op_4473"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<fp16, [1, 1, 12, 128]> var_4469 = transpose(perm = var_4469_perm_0, x = attn_output_97_cast_fp16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 1536]> attn_output_103 = reshape(shape = var_4473, x = var_4469)[name = string("attn_output_103")];
+            tensor<int32, [3]> var_4478 = const()[name = string("op_4478"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_8_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322864832))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324634368))))[name = string("squeeze_8_palettized")];
+            string var_4494_pad_type_0 = const()[name = string("op_4494_pad_type_0"), val = string("valid")];
+            int32 var_4494_groups_0 = const()[name = string("op_4494_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_4494_strides_0 = const()[name = string("op_4494_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_4494_pad_0 = const()[name = string("op_4494_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_4494_dilations_0 = const()[name = string("op_4494_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 1]> var_4479 = transpose(perm = var_4478, x = attn_output_103)[name = string("transpose_3")];
+            tensor<fp16, [1, 1536, 1]> var_4494 = conv(dilations = var_4494_dilations_0, groups = var_4494_groups_0, pad = var_4494_pad_0, pad_type = var_4494_pad_type_0, strides = var_4494_strides_0, weight = squeeze_8_palettized, x = var_4479)[name = string("op_4494")];
+            tensor<int32, [3]> var_4498 = const()[name = string("op_4498"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> attn_output = transpose(perm = var_4498, x = var_4494)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = attn_output)[name = string("hidden_states_cast_fp16")];
+            int32 var_4511 = const()[name = string("op_4511"), val = int32(-1)];
+            fp16 const_194_promoted_to_fp16 = const()[name = string("const_194_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4513_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_194_promoted_to_fp16)[name = string("op_4513_cast_fp16")];
+            bool input_119_interleave_0 = const()[name = string("input_119_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_119_cast_fp16 = concat(axis = var_4511, interleave = input_119_interleave_0, values = (hidden_states_cast_fp16, var_4513_cast_fp16))[name = string("input_119_cast_fp16")];
+            tensor<int32, [1]> normed_69_axes_0 = const()[name = string("normed_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4508_to_fp16 = const()[name = string("op_4508_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_69_cast_fp16 = layer_norm(axes = normed_69_axes_0, epsilon = var_4508_to_fp16, x = input_119_cast_fp16)[name = string("normed_69_cast_fp16")];
+            tensor<int32, [3]> normed_begin_0 = const()[name = string("normed_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_end_0 = const()[name = string("normed_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> normed_end_mask_0 = const()[name = string("normed_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 1536]> normed_cast_fp16 = slice_by_index(begin = normed_begin_0, end = normed_end_0, end_mask = normed_end_mask_0, x = normed_69_cast_fp16)[name = string("normed_cast_fp16")];
+            tensor<fp16, [1536]> const_197_promoted_to_fp16 = const()[name = string("const_197_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324683584)))];
+            tensor<fp16, [1, 1, 1536]> x_141_cast_fp16 = mul(x = normed_cast_fp16, y = const_197_promoted_to_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [3]> var_4538 = const()[name = string("op_4538"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4539 = transpose(perm = var_4538, x = x_141_cast_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1536, 1, 1]> input_121 = expand_dims(axes = input_121_axes_0, x = var_4539)[name = string("input_121")];
+            string input_123_pad_type_0 = const()[name = string("input_123_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_123_strides_0 = const()[name = string("input_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_123_pad_0 = const()[name = string("input_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_123_dilations_0 = const()[name = string("input_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_123_groups_0 = const()[name = string("input_123_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> input_123 = conv(dilations = input_123_dilations_0, groups = input_123_groups_0, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = input_123_strides_0, weight = model_model_layers_18_mlp_gate_proj_weight_palettized, x = input_121)[name = string("input_123")];
+            string b_pad_type_0 = const()[name = string("b_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_strides_0 = const()[name = string("b_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_pad_0 = const()[name = string("b_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_dilations_0 = const()[name = string("b_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_groups_0 = const()[name = string("b_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 1]> b = conv(dilations = b_dilations_0, groups = b_groups_0, pad = b_pad_0, pad_type = b_pad_type_0, strides = b_strides_0, weight = model_model_layers_18_mlp_up_proj_weight_palettized, x = input_121)[name = string("b")];
+            tensor<fp16, [1, 8960, 1, 1]> c = silu(x = input_123)[name = string("c")];
+            tensor<fp16, [1, 8960, 1, 1]> input = mul(x = c, y = b)[name = string("input")];
+            string e_pad_type_0 = const()[name = string("e_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_strides_0 = const()[name = string("e_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_pad_0 = const()[name = string("e_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_dilations_0 = const()[name = string("e_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_groups_0 = const()[name = string("e_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> e = conv(dilations = e_dilations_0, groups = e_groups_0, pad = e_pad_0, pad_type = e_pad_type_0, strides = e_strides_0, weight = model_model_layers_18_mlp_down_proj_weight_palettized, x = input)[name = string("e")];
+            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4561 = squeeze(axes = var_4561_axes_0, x = e)[name = string("op_4561")];
+            tensor<int32, [3]> var_4562 = const()[name = string("op_4562"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1, 1536]> var_4563 = transpose(perm = var_4562, x = var_4561)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 1536]> output_hidden_states = add(x = hidden_states_cast_fp16, y = var_4563)[name = string("op_4565_cast_fp16")];
+            tensor<int32, [1]> position_ids_tmp = identity(x = position_ids)[name = string("position_ids_tmp")];
+        } -> (output_hidden_states);
+    func prefill<ios18>(tensor<fp16, [1, 1, 64, 2048]> causal_mask, tensor<int32, [1]> current_pos, tensor<fp16, [1, 64, 1536]> hidden_states, state<tensor<fp16, [56, 2, 2048, 128]>> model_model_kv_cache_0, tensor<int32, [64]> position_ids) {
+            tensor<fp16, [1536]> model_model_layers_10_self_attn_q_proj_bias = const()[name = string("model_model_layers_10_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_10_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3200))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1772736))))[name = string("model_model_layers_10_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_10_self_attn_k_proj_bias = const()[name = string("model_model_layers_10_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1821952)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_10_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1822528))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2117504))))[name = string("model_model_layers_10_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_10_self_attn_v_proj_bias = const()[name = string("model_model_layers_10_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2125760)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_10_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2126336))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2421312))))[name = string("model_model_layers_10_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_10_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2429568))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12751552))))[name = string("model_model_layers_10_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_10_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13038336))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23360320))))[name = string("model_model_layers_10_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_10_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23647104))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33969088))))[name = string("model_model_layers_10_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_11_self_attn_q_proj_bias = const()[name = string("model_model_layers_11_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34018304)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_11_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34021440))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35790976))))[name = string("model_model_layers_11_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_11_self_attn_k_proj_bias = const()[name = string("model_model_layers_11_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35840192)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_11_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35840768))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36135744))))[name = string("model_model_layers_11_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_11_self_attn_v_proj_bias = const()[name = string("model_model_layers_11_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36144000)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_11_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36144576))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36439552))))[name = string("model_model_layers_11_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_11_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36447808))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46769792))))[name = string("model_model_layers_11_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_11_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47056576))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57378560))))[name = string("model_model_layers_11_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_11_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57665344))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67987328))))[name = string("model_model_layers_11_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_12_self_attn_q_proj_bias = const()[name = string("model_model_layers_12_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68036544)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_12_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68039680))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69809216))))[name = string("model_model_layers_12_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_12_self_attn_k_proj_bias = const()[name = string("model_model_layers_12_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69858432)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_12_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69859008))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70153984))))[name = string("model_model_layers_12_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_12_self_attn_v_proj_bias = const()[name = string("model_model_layers_12_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70162240)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_12_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70162816))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70457792))))[name = string("model_model_layers_12_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_12_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70466048))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80788032))))[name = string("model_model_layers_12_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_12_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81074816))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91396800))))[name = string("model_model_layers_12_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_12_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91683584))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102005568))))[name = string("model_model_layers_12_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_13_self_attn_q_proj_bias = const()[name = string("model_model_layers_13_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102054784)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_13_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102057920))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103827456))))[name = string("model_model_layers_13_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_13_self_attn_k_proj_bias = const()[name = string("model_model_layers_13_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103876672)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_13_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103877248))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104172224))))[name = string("model_model_layers_13_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_13_self_attn_v_proj_bias = const()[name = string("model_model_layers_13_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104180480)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_13_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104181056))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104476032))))[name = string("model_model_layers_13_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_13_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104484288))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114806272))))[name = string("model_model_layers_13_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_13_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115093056))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125415040))))[name = string("model_model_layers_13_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_13_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125701824))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136023808))))[name = string("model_model_layers_13_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_14_self_attn_q_proj_bias = const()[name = string("model_model_layers_14_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136073024)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_14_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136076160))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137845696))))[name = string("model_model_layers_14_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_14_self_attn_k_proj_bias = const()[name = string("model_model_layers_14_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137894912)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_14_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137895488))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138190464))))[name = string("model_model_layers_14_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_14_self_attn_v_proj_bias = const()[name = string("model_model_layers_14_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138198720)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_14_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138199296))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138494272))))[name = string("model_model_layers_14_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_14_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138502528))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148824512))))[name = string("model_model_layers_14_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_14_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149111296))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159433280))))[name = string("model_model_layers_14_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_14_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159720064))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170042048))))[name = string("model_model_layers_14_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_15_self_attn_q_proj_bias = const()[name = string("model_model_layers_15_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170091264)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_15_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170094400))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171863936))))[name = string("model_model_layers_15_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_15_self_attn_k_proj_bias = const()[name = string("model_model_layers_15_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171913152)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_15_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171913728))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172208704))))[name = string("model_model_layers_15_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_15_self_attn_v_proj_bias = const()[name = string("model_model_layers_15_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172216960)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_15_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172217536))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172512512))))[name = string("model_model_layers_15_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_15_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172520768))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182842752))))[name = string("model_model_layers_15_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_15_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183129536))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193451520))))[name = string("model_model_layers_15_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_15_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193738304))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204060288))))[name = string("model_model_layers_15_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_16_self_attn_q_proj_bias = const()[name = string("model_model_layers_16_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204109504)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_16_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204112640))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205882176))))[name = string("model_model_layers_16_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_16_self_attn_k_proj_bias = const()[name = string("model_model_layers_16_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205931392)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_16_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205931968))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206226944))))[name = string("model_model_layers_16_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_16_self_attn_v_proj_bias = const()[name = string("model_model_layers_16_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206235200)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_16_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206235776))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206530752))))[name = string("model_model_layers_16_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_16_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206539008))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216860992))))[name = string("model_model_layers_16_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_16_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217147776))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227469760))))[name = string("model_model_layers_16_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_16_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227756544))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238078528))))[name = string("model_model_layers_16_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_17_self_attn_q_proj_bias = const()[name = string("model_model_layers_17_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238127744)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_17_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238130880))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239900416))))[name = string("model_model_layers_17_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_17_self_attn_k_proj_bias = const()[name = string("model_model_layers_17_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239949632)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_17_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239950208))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240245184))))[name = string("model_model_layers_17_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_17_self_attn_v_proj_bias = const()[name = string("model_model_layers_17_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240253440)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_17_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240254016))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240548992))))[name = string("model_model_layers_17_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_17_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240557248))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250879232))))[name = string("model_model_layers_17_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_17_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251166016))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261488000))))[name = string("model_model_layers_17_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_17_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261774784))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272096768))))[name = string("model_model_layers_17_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [1536]> model_model_layers_18_self_attn_q_proj_bias = const()[name = string("model_model_layers_18_self_attn_q_proj_bias"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272145984)))];
+            tensor<fp16, [1536, 1536, 1, 1]> model_model_layers_18_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272149120))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273918656))))[name = string("model_model_layers_18_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_18_self_attn_k_proj_bias = const()[name = string("model_model_layers_18_self_attn_k_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273967872)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_18_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273968448))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274263424))))[name = string("model_model_layers_18_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256]> model_model_layers_18_self_attn_v_proj_bias = const()[name = string("model_model_layers_18_self_attn_v_proj_bias"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274271680)))];
+            tensor<fp16, [256, 1536, 1, 1]> model_model_layers_18_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274272256))), lut = tensor<fp16, [64, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274567232))))[name = string("model_model_layers_18_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_18_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274575488))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284897472))))[name = string("model_model_layers_18_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [8960, 1536, 1, 1]> model_model_layers_18_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [8960, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285184256))), lut = tensor<fp16, [2240, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295506240))))[name = string("model_model_layers_18_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 8960, 1, 1]> model_model_layers_18_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 8960, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295793024))), lut = tensor<fp16, [384, 1, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306115008))))[name = string("model_model_layers_18_mlp_down_proj_weight_palettized")];
+            int32 var_390_batch_dims_0 = const()[name = string("op_390_batch_dims_0"), val = int32(0)];
+            bool var_390_validate_indices_0 = const()[name = string("op_390_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [1, 4096, 128]> var_382_to_fp16 = const()[name = string("op_382_to_fp16"), val = tensor<fp16, [1, 4096, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307212864)))];
+            string position_ids_to_int16_dtype_0 = const()[name = string("position_ids_to_int16_dtype_0"), val = string("int16")];
+            string cast_78_dtype_0 = const()[name = string("cast_78_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [64]> position_ids_to_int16 = cast(dtype = position_ids_to_int16_dtype_0, x = position_ids)[name = string("cast_5")];
+            tensor<int32, [64]> cast_78 = cast(dtype = cast_78_dtype_0, x = position_ids_to_int16)[name = string("cast_4")];
+            tensor<bool, [64]> greater_equal_0 = greater_equal(x = cast_78, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_72 = const()[name = string("slice_by_index_72"), val = int32(4096)];
+            tensor<int32, [64]> add_0 = add(x = cast_78, y = slice_by_index_72)[name = string("add_0")];
+            tensor<int32, [64]> select_0 = select(a = cast_78, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [64]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_3")];
+            tensor<int32, [64]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_2")];
+            tensor<bool, [64]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(4096)];
+            tensor<int32, [64]> add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")];
+            tensor<int32, [64]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 op_390_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("op_390_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 64, 128]> op_390_cast_fp16_cast_uint16_cast_uint16 = gather(axis = op_390_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = var_390_batch_dims_0, indices = select_0_1, validate_indices = var_390_validate_indices_0, x = var_382_to_fp16)[name = string("op_390_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [4]> var_394 = const()[name = string("op_394"), val = tensor<int32, [4]>([1, 64, 1, 128])];
+            tensor<fp16, [1, 64, 1, 128]> cos_1_cast_fp16 = reshape(shape = var_394, x = op_390_cast_fp16_cast_uint16_cast_uint16)[name = string("cos_1_cast_fp16")];
+            int32 var_404_axis_0 = const()[name = string("op_404_axis_0"), val = int32(1)];
+            int32 var_404_batch_dims_0 = const()[name = string("op_404_batch_dims_0"), val = int32(0)];
+            bool var_404_validate_indices_0 = const()[name = string("op_404_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [1, 4096, 128]> var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = tensor<fp16, [1, 4096, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306164224)))];
+            string position_ids_to_uint16_dtype_0 = const()[name = string("position_ids_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [64]> position_ids_to_uint16 = cast(dtype = position_ids_to_uint16_dtype_0, x = position_ids)[name = string("cast_1")];
+            tensor<fp16, [1, 64, 128]> var_404_cast_fp16_cast_uint16 = gather(axis = var_404_axis_0, batch_dims = var_404_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_404_validate_indices_0, x = var_396_to_fp16)[name = string("op_404_cast_fp16_cast_uint16")];
+            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 64, 1, 128])];
+            tensor<fp16, [1, 64, 1, 128]> sin_1_cast_fp16 = reshape(shape = var_408, x = var_404_cast_fp16_cast_uint16)[name = string("sin_1_cast_fp16")];
+            int32 var_429 = const()[name = string("op_429"), val = int32(-1)];
+            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_431_cast_fp16 = mul(x = hidden_states, y = const_1_promoted_to_fp16)[name = string("op_431_cast_fp16")];
+            bool input_1_interleave_0 = const()[name = string("input_1_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_1_cast_fp16 = concat(axis = var_429, interleave = input_1_interleave_0, values = (hidden_states, var_431_cast_fp16))[name = string("input_1_cast_fp16")];
+            tensor<int32, [1]> normed_1_axes_0 = const()[name = string("normed_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_426_to_fp16 = const()[name = string("op_426_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_1_cast_fp16 = layer_norm(axes = normed_1_axes_0, epsilon = var_426_to_fp16, x = input_1_cast_fp16)[name = string("normed_1_cast_fp16")];
+            tensor<int32, [3]> normed_3_begin_0 = const()[name = string("normed_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_3_end_0 = const()[name = string("normed_3_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_3_end_mask_0 = const()[name = string("normed_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_3_cast_fp16 = slice_by_index(begin = normed_3_begin_0, end = normed_3_end_0, end_mask = normed_3_end_mask_0, x = normed_1_cast_fp16)[name = string("normed_3_cast_fp16")];
+            tensor<fp16, [1536]> const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308261504)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_3_cast_fp16 = mul(x = normed_3_cast_fp16, y = const_4_promoted_to_fp16)[name = string("hidden_states_3_cast_fp16")];
+            tensor<int32, [3]> var_454 = const()[name = string("op_454"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_457_axes_0 = const()[name = string("op_457_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_455_cast_fp16 = transpose(perm = var_454, x = hidden_states_3_cast_fp16)[name = string("transpose_82")];
+            tensor<fp16, [1, 1536, 1, 64]> var_457_cast_fp16 = expand_dims(axes = var_457_axes_0, x = var_455_cast_fp16)[name = string("op_457_cast_fp16")];
+            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_1 = conv(bias = model_model_layers_10_self_attn_q_proj_bias, dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = model_model_layers_10_self_attn_q_proj_weight_palettized, x = var_457_cast_fp16)[name = string("query_states_1")];
+            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_1 = conv(bias = model_model_layers_10_self_attn_k_proj_bias, dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = model_model_layers_10_self_attn_k_proj_weight_palettized, x = var_457_cast_fp16)[name = string("key_states_1")];
+            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_1 = conv(bias = model_model_layers_10_self_attn_v_proj_bias, dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = model_model_layers_10_self_attn_v_proj_weight_palettized, x = var_457_cast_fp16)[name = string("value_states_1")];
+            tensor<int32, [4]> var_499 = const()[name = string("op_499"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_500 = reshape(shape = var_499, x = query_states_1)[name = string("op_500")];
+            tensor<int32, [4]> var_505 = const()[name = string("op_505"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_510 = const()[name = string("op_510"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_511 = reshape(shape = var_510, x = key_states_1)[name = string("op_511")];
+            tensor<int32, [4]> var_516 = const()[name = string("op_516"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_521 = const()[name = string("op_521"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_522 = reshape(shape = var_521, x = value_states_1)[name = string("op_522")];
+            tensor<int32, [4]> var_527 = const()[name = string("op_527"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_533 = const()[name = string("op_533"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_539 = const()[name = string("op_539"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [1, 12, 64, 128]> q_1 = transpose(perm = var_505, x = var_500)[name = string("transpose_80")];
+            tensor<fp16, [1, 1, 64, 128]> cos_5 = transpose(perm = var_533, x = cos_1_cast_fp16)[name = string("transpose_81")];
+            tensor<fp16, [1, 12, 64, 128]> var_541 = mul(x = q_1, y = cos_5)[name = string("op_541")];
+            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_1 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1)[name = string("x1_1")];
+            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_1 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1)[name = string("x2_1")];
+            fp16 const_8_promoted = const()[name = string("const_8_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_562 = mul(x = x2_1, y = const_8_promoted)[name = string("op_562")];
+            int32 var_564 = const()[name = string("op_564"), val = int32(-1)];
+            bool var_565_interleave_0 = const()[name = string("op_565_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_565 = concat(axis = var_564, interleave = var_565_interleave_0, values = (var_562, x1_1))[name = string("op_565")];
+            tensor<fp16, [1, 1, 64, 128]> sin_5 = transpose(perm = var_539, x = sin_1_cast_fp16)[name = string("transpose_79")];
+            tensor<fp16, [1, 12, 64, 128]> var_566 = mul(x = var_565, y = sin_5)[name = string("op_566")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_3 = add(x = var_541, y = var_566)[name = string("query_states_3")];
+            tensor<fp16, [1, 2, 64, 128]> k_1 = transpose(perm = var_516, x = var_511)[name = string("transpose_78")];
+            tensor<fp16, [1, 2, 64, 128]> var_569 = mul(x = k_1, y = cos_5)[name = string("op_569")];
+            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_3 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1)[name = string("x1_3")];
+            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_3 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1)[name = string("x2_3")];
+            fp16 const_11_promoted = const()[name = string("const_11_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_590 = mul(x = x2_3, y = const_11_promoted)[name = string("op_590")];
+            int32 var_592 = const()[name = string("op_592"), val = int32(-1)];
+            bool var_593_interleave_0 = const()[name = string("op_593_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_593 = concat(axis = var_592, interleave = var_593_interleave_0, values = (var_590, x1_3))[name = string("op_593")];
+            tensor<fp16, [1, 2, 64, 128]> var_594 = mul(x = var_593, y = sin_5)[name = string("op_594")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_3 = add(x = var_569, y = var_594)[name = string("key_states_3")];
+            tensor<int32, [1]> seq_length_1 = const()[name = string("seq_length_1"), val = tensor<int32, [1]>([64])];
+            tensor<int32, [1]> var_616 = add(x = current_pos, y = seq_length_1)[name = string("op_616")];
+            tensor<fp16, [56, 2, 2048, 128]> read_state_0 = read_state(input = model_model_kv_cache_0)[name = string("read_state_0")];
+            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([10])];
+            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([11])];
+            int32 concat_2_axis_0 = const()[name = string("concat_2_axis_0"), val = int32(0)];
+            bool concat_2_interleave_0 = const()[name = string("concat_2_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_2 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = (expand_dims_0, expand_dims_1, current_pos, expand_dims_3))[name = string("concat_2")];
+            tensor<int32, [1]> concat_3_values1_0 = const()[name = string("concat_3_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_3_values3_0 = const()[name = string("concat_3_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_3_axis_0 = const()[name = string("concat_3_axis_0"), val = int32(0)];
+            bool concat_3_interleave_0 = const()[name = string("concat_3_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (expand_dims_4, concat_3_values1_0, var_616, concat_3_values3_0))[name = string("concat_3")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_1_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_2, begin_mask = model_model_kv_cache_0_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = model_model_kv_cache_0_internal_tensor_assign_1_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_1_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_1_stride_0, update = key_states_3, x = read_state_0)[name = string("model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_1_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_18_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_18 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_18")];
+            tensor<int32, [1]> expand_dims_6 = const()[name = string("expand_dims_6"), val = tensor<int32, [1]>([38])];
+            tensor<int32, [1]> expand_dims_7 = const()[name = string("expand_dims_7"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_9 = const()[name = string("expand_dims_9"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_10 = const()[name = string("expand_dims_10"), val = tensor<int32, [1]>([39])];
+            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
+            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_6, expand_dims_7, current_pos, expand_dims_9))[name = string("concat_6")];
+            tensor<int32, [1]> concat_7_values1_0 = const()[name = string("concat_7_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_7_values3_0 = const()[name = string("concat_7_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_7_axis_0 = const()[name = string("concat_7_axis_0"), val = int32(0)];
+            bool concat_7_interleave_0 = const()[name = string("concat_7_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_7 = concat(axis = concat_7_axis_0, interleave = concat_7_interleave_0, values = (expand_dims_10, concat_7_values1_0, var_616, concat_7_values3_0))[name = string("concat_7")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_2_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_3 = transpose(perm = var_527, x = var_522)[name = string("transpose_77")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_6, begin_mask = model_model_kv_cache_0_internal_tensor_assign_2_begin_mask_0, end = concat_7, end_mask = model_model_kv_cache_0_internal_tensor_assign_2_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_2_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_2_stride_0, update = value_states_3, x = coreml_update_state_18)[name = string("model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_2_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_19_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_19 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_19")];
+            tensor<int32, [4]> var_665_begin_0 = const()[name = string("op_665_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> var_665_end_0 = const()[name = string("op_665_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
+            tensor<bool, [4]> var_665_end_mask_0 = const()[name = string("op_665_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_665_cast_fp16 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = coreml_update_state_19)[name = string("op_665_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_1_axes_0 = const()[name = string("K_layer_cache_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_1_cast_fp16 = squeeze(axes = K_layer_cache_1_axes_0, x = var_665_cast_fp16)[name = string("K_layer_cache_1_cast_fp16")];
+            tensor<int32, [4]> var_672_begin_0 = const()[name = string("op_672_begin_0"), val = tensor<int32, [4]>([38, 0, 0, 0])];
+            tensor<int32, [4]> var_672_end_0 = const()[name = string("op_672_end_0"), val = tensor<int32, [4]>([39, 2, 2048, 128])];
+            tensor<bool, [4]> var_672_end_mask_0 = const()[name = string("op_672_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_672_cast_fp16 = slice_by_index(begin = var_672_begin_0, end = var_672_end_0, end_mask = var_672_end_mask_0, x = coreml_update_state_19)[name = string("op_672_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_1_axes_0 = const()[name = string("V_layer_cache_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_1_cast_fp16 = squeeze(axes = V_layer_cache_1_axes_0, x = var_672_cast_fp16)[name = string("V_layer_cache_1_cast_fp16")];
+            tensor<int32, [1]> x_3_axes_0 = const()[name = string("x_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_3_cast_fp16 = expand_dims(axes = x_3_axes_0, x = K_layer_cache_1_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_701 = const()[name = string("op_701"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_5_cast_fp16 = tile(reps = var_701, x = x_3_cast_fp16)[name = string("x_5_cast_fp16")];
+            tensor<int32, [4]> var_713 = const()[name = string("op_713"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_7_cast_fp16 = reshape(shape = var_713, x = x_5_cast_fp16)[name = string("key_states_7_cast_fp16")];
+            tensor<int32, [1]> x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_9_cast_fp16 = expand_dims(axes = x_9_axes_0, x = V_layer_cache_1_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_721 = const()[name = string("op_721"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_11_cast_fp16 = tile(reps = var_721, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            bool var_756_transpose_x_1 = const()[name = string("op_756_transpose_x_1"), val = bool(false)];
+            bool var_756_transpose_y_1 = const()[name = string("op_756_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_756_cast_fp16 = matmul(transpose_x = var_756_transpose_x_1, transpose_y = var_756_transpose_y_1, x = query_states_3, y = key_states_7_cast_fp16)[name = string("op_756_cast_fp16")];
+            fp16 var_757_to_fp16 = const()[name = string("op_757_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_1_cast_fp16 = mul(x = var_756_cast_fp16, y = var_757_to_fp16)[name = string("attn_logits_1_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_3_cast_fp16 = add(x = attn_logits_1_cast_fp16, y = causal_mask)[name = string("attn_logits_3_cast_fp16")];
+            int32 var_784 = const()[name = string("op_784"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_786_cast_fp16 = softmax(axis = var_784, x = attn_logits_3_cast_fp16)[name = string("op_786_cast_fp16")];
+            tensor<int32, [3]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_0_cast_fp16 = reshape(shape = concat_12, x = var_786_cast_fp16)[name = string("reshape_0_cast_fp16")];
+            tensor<int32, [3]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_1_cast_fp16 = reshape(shape = concat_13, x = x_11_cast_fp16)[name = string("reshape_1_cast_fp16")];
+            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
+            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = reshape_0_cast_fp16, y = reshape_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
+            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_2_cast_fp16 = reshape(shape = concat_17, x = matmul_0_cast_fp16)[name = string("reshape_2_cast_fp16")];
+            tensor<int32, [4]> var_813_perm_0 = const()[name = string("op_813_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_832 = const()[name = string("op_832"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_813 = transpose(perm = var_813_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_76")];
+            tensor<fp16, [1, 64, 1536]> attn_output_5 = reshape(shape = var_832, x = var_813)[name = string("attn_output_5")];
+            tensor<int32, [3]> var_837 = const()[name = string("op_837"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_0_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308264640))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310034176))))[name = string("squeeze_0_palettized")];
+            string var_853_pad_type_0 = const()[name = string("op_853_pad_type_0"), val = string("valid")];
+            int32 var_853_groups_0 = const()[name = string("op_853_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_853_strides_0 = const()[name = string("op_853_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_853_pad_0 = const()[name = string("op_853_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_853_dilations_0 = const()[name = string("op_853_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_838 = transpose(perm = var_837, x = attn_output_5)[name = string("transpose_75")];
+            tensor<fp16, [1, 1536, 64]> var_853 = conv(dilations = var_853_dilations_0, groups = var_853_groups_0, pad = var_853_pad_0, pad_type = var_853_pad_type_0, strides = var_853_strides_0, weight = squeeze_0_palettized, x = var_838)[name = string("op_853")];
+            tensor<int32, [3]> var_857 = const()[name = string("op_857"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_9 = transpose(perm = var_857, x = var_853)[name = string("transpose_74")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_5_cast_fp16 = add(x = hidden_states, y = attn_output_9)[name = string("hidden_states_5_cast_fp16")];
+            int32 var_870 = const()[name = string("op_870"), val = int32(-1)];
+            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_872_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_872_cast_fp16")];
+            bool input_7_interleave_0 = const()[name = string("input_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_7_cast_fp16 = concat(axis = var_870, interleave = input_7_interleave_0, values = (hidden_states_5_cast_fp16, var_872_cast_fp16))[name = string("input_7_cast_fp16")];
+            tensor<int32, [1]> normed_5_axes_0 = const()[name = string("normed_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_867_to_fp16 = const()[name = string("op_867_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_5_cast_fp16 = layer_norm(axes = normed_5_axes_0, epsilon = var_867_to_fp16, x = input_7_cast_fp16)[name = string("normed_5_cast_fp16")];
+            tensor<int32, [3]> normed_7_begin_0 = const()[name = string("normed_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_7_end_0 = const()[name = string("normed_7_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_7_end_mask_0 = const()[name = string("normed_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_7_cast_fp16 = slice_by_index(begin = normed_7_begin_0, end = normed_7_end_0, end_mask = normed_7_end_mask_0, x = normed_5_cast_fp16)[name = string("normed_7_cast_fp16")];
+            tensor<fp16, [1536]> const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310083392)))];
+            tensor<fp16, [1, 64, 1536]> x_13_cast_fp16 = mul(x = normed_7_cast_fp16, y = const_26_promoted_to_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [3]> var_897 = const()[name = string("op_897"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_898 = transpose(perm = var_897, x = x_13_cast_fp16)[name = string("transpose_73")];
+            tensor<fp16, [1, 1536, 1, 64]> input_9 = expand_dims(axes = input_9_axes_0, x = var_898)[name = string("input_9")];
+            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_11 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = model_model_layers_10_mlp_gate_proj_weight_palettized, x = input_9)[name = string("input_11")];
+            string b_1_pad_type_0 = const()[name = string("b_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_1_strides_0 = const()[name = string("b_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_1_pad_0 = const()[name = string("b_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_1_dilations_0 = const()[name = string("b_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_1_groups_0 = const()[name = string("b_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_1 = conv(dilations = b_1_dilations_0, groups = b_1_groups_0, pad = b_1_pad_0, pad_type = b_1_pad_type_0, strides = b_1_strides_0, weight = model_model_layers_10_mlp_up_proj_weight_palettized, x = input_9)[name = string("b_1")];
+            tensor<fp16, [1, 8960, 1, 64]> c_1 = silu(x = input_11)[name = string("c_1")];
+            tensor<fp16, [1, 8960, 1, 64]> input_13 = mul(x = c_1, y = b_1)[name = string("input_13")];
+            string e_1_pad_type_0 = const()[name = string("e_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_1_strides_0 = const()[name = string("e_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_1_pad_0 = const()[name = string("e_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_1_dilations_0 = const()[name = string("e_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_1_groups_0 = const()[name = string("e_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_1 = conv(dilations = e_1_dilations_0, groups = e_1_groups_0, pad = e_1_pad_0, pad_type = e_1_pad_type_0, strides = e_1_strides_0, weight = model_model_layers_10_mlp_down_proj_weight_palettized, x = input_13)[name = string("e_1")];
+            tensor<int32, [1]> var_920_axes_0 = const()[name = string("op_920_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_920 = squeeze(axes = var_920_axes_0, x = e_1)[name = string("op_920")];
+            tensor<int32, [3]> var_921 = const()[name = string("op_921"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_922 = transpose(perm = var_921, x = var_920)[name = string("transpose_72")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_7_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = var_922)[name = string("hidden_states_7_cast_fp16")];
+            int32 var_934 = const()[name = string("op_934"), val = int32(-1)];
+            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_936_cast_fp16 = mul(x = hidden_states_7_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_936_cast_fp16")];
+            bool input_15_interleave_0 = const()[name = string("input_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_15_cast_fp16 = concat(axis = var_934, interleave = input_15_interleave_0, values = (hidden_states_7_cast_fp16, var_936_cast_fp16))[name = string("input_15_cast_fp16")];
+            tensor<int32, [1]> normed_9_axes_0 = const()[name = string("normed_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_931_to_fp16 = const()[name = string("op_931_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_9_cast_fp16 = layer_norm(axes = normed_9_axes_0, epsilon = var_931_to_fp16, x = input_15_cast_fp16)[name = string("normed_9_cast_fp16")];
+            tensor<int32, [3]> normed_11_begin_0 = const()[name = string("normed_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_11_end_0 = const()[name = string("normed_11_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_11_end_mask_0 = const()[name = string("normed_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_11_cast_fp16 = slice_by_index(begin = normed_11_begin_0, end = normed_11_end_0, end_mask = normed_11_end_mask_0, x = normed_9_cast_fp16)[name = string("normed_11_cast_fp16")];
+            tensor<fp16, [1536]> const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310086528)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_9_cast_fp16 = mul(x = normed_11_cast_fp16, y = const_30_promoted_to_fp16)[name = string("hidden_states_9_cast_fp16")];
+            tensor<int32, [3]> var_959 = const()[name = string("op_959"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_962_axes_0 = const()[name = string("op_962_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_960_cast_fp16 = transpose(perm = var_959, x = hidden_states_9_cast_fp16)[name = string("transpose_71")];
+            tensor<fp16, [1, 1536, 1, 64]> var_962_cast_fp16 = expand_dims(axes = var_962_axes_0, x = var_960_cast_fp16)[name = string("op_962_cast_fp16")];
+            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_7 = conv(bias = model_model_layers_11_self_attn_q_proj_bias, dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = model_model_layers_11_self_attn_q_proj_weight_palettized, x = var_962_cast_fp16)[name = string("query_states_7")];
+            string key_states_9_pad_type_0 = const()[name = string("key_states_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_9_strides_0 = const()[name = string("key_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_9_pad_0 = const()[name = string("key_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_9_dilations_0 = const()[name = string("key_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_9_groups_0 = const()[name = string("key_states_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_9 = conv(bias = model_model_layers_11_self_attn_k_proj_bias, dilations = key_states_9_dilations_0, groups = key_states_9_groups_0, pad = key_states_9_pad_0, pad_type = key_states_9_pad_type_0, strides = key_states_9_strides_0, weight = model_model_layers_11_self_attn_k_proj_weight_palettized, x = var_962_cast_fp16)[name = string("key_states_9")];
+            string value_states_9_pad_type_0 = const()[name = string("value_states_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_9_strides_0 = const()[name = string("value_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_9_pad_0 = const()[name = string("value_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_9_dilations_0 = const()[name = string("value_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_9_groups_0 = const()[name = string("value_states_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_9 = conv(bias = model_model_layers_11_self_attn_v_proj_bias, dilations = value_states_9_dilations_0, groups = value_states_9_groups_0, pad = value_states_9_pad_0, pad_type = value_states_9_pad_type_0, strides = value_states_9_strides_0, weight = model_model_layers_11_self_attn_v_proj_weight_palettized, x = var_962_cast_fp16)[name = string("value_states_9")];
+            tensor<int32, [4]> var_1004 = const()[name = string("op_1004"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_1005 = reshape(shape = var_1004, x = query_states_7)[name = string("op_1005")];
+            tensor<int32, [4]> var_1010 = const()[name = string("op_1010"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_1015 = const()[name = string("op_1015"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_1016 = reshape(shape = var_1015, x = key_states_9)[name = string("op_1016")];
+            tensor<int32, [4]> var_1021 = const()[name = string("op_1021"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_1026 = const()[name = string("op_1026"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_1027 = reshape(shape = var_1026, x = value_states_9)[name = string("op_1027")];
+            tensor<int32, [4]> var_1032 = const()[name = string("op_1032"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_5 = transpose(perm = var_1010, x = var_1005)[name = string("transpose_70")];
+            tensor<fp16, [1, 12, 64, 128]> var_1046 = mul(x = q_5, y = cos_5)[name = string("op_1046")];
+            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_5 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_5)[name = string("x1_5")];
+            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_5 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_5)[name = string("x2_5")];
+            fp16 const_34_promoted = const()[name = string("const_34_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_1067 = mul(x = x2_5, y = const_34_promoted)[name = string("op_1067")];
+            int32 var_1069 = const()[name = string("op_1069"), val = int32(-1)];
+            bool var_1070_interleave_0 = const()[name = string("op_1070_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_1070 = concat(axis = var_1069, interleave = var_1070_interleave_0, values = (var_1067, x1_5))[name = string("op_1070")];
+            tensor<fp16, [1, 12, 64, 128]> var_1071 = mul(x = var_1070, y = sin_5)[name = string("op_1071")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_9 = add(x = var_1046, y = var_1071)[name = string("query_states_9")];
+            tensor<fp16, [1, 2, 64, 128]> k_5 = transpose(perm = var_1021, x = var_1016)[name = string("transpose_69")];
+            tensor<fp16, [1, 2, 64, 128]> var_1074 = mul(x = k_5, y = cos_5)[name = string("op_1074")];
+            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_7 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_5)[name = string("x1_7")];
+            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_7 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_5)[name = string("x2_7")];
+            fp16 const_37_promoted = const()[name = string("const_37_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_1095 = mul(x = x2_7, y = const_37_promoted)[name = string("op_1095")];
+            int32 var_1097 = const()[name = string("op_1097"), val = int32(-1)];
+            bool var_1098_interleave_0 = const()[name = string("op_1098_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_1098 = concat(axis = var_1097, interleave = var_1098_interleave_0, values = (var_1095, x1_7))[name = string("op_1098")];
+            tensor<fp16, [1, 2, 64, 128]> var_1099 = mul(x = var_1098, y = sin_5)[name = string("op_1099")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_11 = add(x = var_1074, y = var_1099)[name = string("key_states_11")];
+            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([11])];
+            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([12])];
+            int32 concat_20_axis_0 = const()[name = string("concat_20_axis_0"), val = int32(0)];
+            bool concat_20_interleave_0 = const()[name = string("concat_20_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_20 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (expand_dims_12, expand_dims_13, current_pos, expand_dims_15))[name = string("concat_20")];
+            tensor<int32, [1]> concat_21_values1_0 = const()[name = string("concat_21_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_21_values3_0 = const()[name = string("concat_21_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_21_axis_0 = const()[name = string("concat_21_axis_0"), val = int32(0)];
+            bool concat_21_interleave_0 = const()[name = string("concat_21_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_21 = concat(axis = concat_21_axis_0, interleave = concat_21_interleave_0, values = (expand_dims_16, concat_21_values1_0, var_616, concat_21_values3_0))[name = string("concat_21")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_3_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_20, begin_mask = model_model_kv_cache_0_internal_tensor_assign_3_begin_mask_0, end = concat_21, end_mask = model_model_kv_cache_0_internal_tensor_assign_3_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_3_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_3_stride_0, update = key_states_11, x = coreml_update_state_19)[name = string("model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_3_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_20_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_20 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_20")];
+            tensor<int32, [1]> expand_dims_18 = const()[name = string("expand_dims_18"), val = tensor<int32, [1]>([39])];
+            tensor<int32, [1]> expand_dims_19 = const()[name = string("expand_dims_19"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_21 = const()[name = string("expand_dims_21"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_22 = const()[name = string("expand_dims_22"), val = tensor<int32, [1]>([40])];
+            int32 concat_24_axis_0 = const()[name = string("concat_24_axis_0"), val = int32(0)];
+            bool concat_24_interleave_0 = const()[name = string("concat_24_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_24 = concat(axis = concat_24_axis_0, interleave = concat_24_interleave_0, values = (expand_dims_18, expand_dims_19, current_pos, expand_dims_21))[name = string("concat_24")];
+            tensor<int32, [1]> concat_25_values1_0 = const()[name = string("concat_25_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_25_values3_0 = const()[name = string("concat_25_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_25_axis_0 = const()[name = string("concat_25_axis_0"), val = int32(0)];
+            bool concat_25_interleave_0 = const()[name = string("concat_25_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_25 = concat(axis = concat_25_axis_0, interleave = concat_25_interleave_0, values = (expand_dims_22, concat_25_values1_0, var_616, concat_25_values3_0))[name = string("concat_25")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_4_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_11 = transpose(perm = var_1032, x = var_1027)[name = string("transpose_68")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_24, begin_mask = model_model_kv_cache_0_internal_tensor_assign_4_begin_mask_0, end = concat_25, end_mask = model_model_kv_cache_0_internal_tensor_assign_4_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_4_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_4_stride_0, update = value_states_11, x = coreml_update_state_20)[name = string("model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_4_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_21_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_21 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_21")];
+            tensor<int32, [4]> var_1170_begin_0 = const()[name = string("op_1170_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> var_1170_end_0 = const()[name = string("op_1170_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
+            tensor<bool, [4]> var_1170_end_mask_0 = const()[name = string("op_1170_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1170_cast_fp16 = slice_by_index(begin = var_1170_begin_0, end = var_1170_end_0, end_mask = var_1170_end_mask_0, x = coreml_update_state_21)[name = string("op_1170_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_3_axes_0 = const()[name = string("K_layer_cache_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_3_cast_fp16 = squeeze(axes = K_layer_cache_3_axes_0, x = var_1170_cast_fp16)[name = string("K_layer_cache_3_cast_fp16")];
+            tensor<int32, [4]> var_1177_begin_0 = const()[name = string("op_1177_begin_0"), val = tensor<int32, [4]>([39, 0, 0, 0])];
+            tensor<int32, [4]> var_1177_end_0 = const()[name = string("op_1177_end_0"), val = tensor<int32, [4]>([40, 2, 2048, 128])];
+            tensor<bool, [4]> var_1177_end_mask_0 = const()[name = string("op_1177_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1177_cast_fp16 = slice_by_index(begin = var_1177_begin_0, end = var_1177_end_0, end_mask = var_1177_end_mask_0, x = coreml_update_state_21)[name = string("op_1177_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_3_axes_0 = const()[name = string("V_layer_cache_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_3_cast_fp16 = squeeze(axes = V_layer_cache_3_axes_0, x = var_1177_cast_fp16)[name = string("V_layer_cache_3_cast_fp16")];
+            tensor<int32, [1]> x_19_axes_0 = const()[name = string("x_19_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_19_cast_fp16 = expand_dims(axes = x_19_axes_0, x = K_layer_cache_3_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [4]> var_1206 = const()[name = string("op_1206"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_21_cast_fp16 = tile(reps = var_1206, x = x_19_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [4]> var_1218 = const()[name = string("op_1218"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_15_cast_fp16 = reshape(shape = var_1218, x = x_21_cast_fp16)[name = string("key_states_15_cast_fp16")];
+            tensor<int32, [1]> x_25_axes_0 = const()[name = string("x_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_25_cast_fp16 = expand_dims(axes = x_25_axes_0, x = V_layer_cache_3_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [4]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_27_cast_fp16 = tile(reps = var_1226, x = x_25_cast_fp16)[name = string("x_27_cast_fp16")];
+            bool var_1261_transpose_x_1 = const()[name = string("op_1261_transpose_x_1"), val = bool(false)];
+            bool var_1261_transpose_y_1 = const()[name = string("op_1261_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_1261_cast_fp16 = matmul(transpose_x = var_1261_transpose_x_1, transpose_y = var_1261_transpose_y_1, x = query_states_9, y = key_states_15_cast_fp16)[name = string("op_1261_cast_fp16")];
+            fp16 var_1262_to_fp16 = const()[name = string("op_1262_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_5_cast_fp16 = mul(x = var_1261_cast_fp16, y = var_1262_to_fp16)[name = string("attn_logits_5_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_7_cast_fp16 = add(x = attn_logits_5_cast_fp16, y = causal_mask)[name = string("attn_logits_7_cast_fp16")];
+            int32 var_1289 = const()[name = string("op_1289"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_1291_cast_fp16 = softmax(axis = var_1289, x = attn_logits_7_cast_fp16)[name = string("op_1291_cast_fp16")];
+            tensor<int32, [3]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_3_cast_fp16 = reshape(shape = concat_30, x = var_1291_cast_fp16)[name = string("reshape_3_cast_fp16")];
+            tensor<int32, [3]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_4_cast_fp16 = reshape(shape = concat_31, x = x_27_cast_fp16)[name = string("reshape_4_cast_fp16")];
+            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
+            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = reshape_3_cast_fp16, y = reshape_4_cast_fp16)[name = string("matmul_1_cast_fp16")];
+            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_5_cast_fp16 = reshape(shape = concat_35, x = matmul_1_cast_fp16)[name = string("reshape_5_cast_fp16")];
+            tensor<int32, [4]> var_1318_perm_0 = const()[name = string("op_1318_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1337 = const()[name = string("op_1337"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_1318 = transpose(perm = var_1318_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_67")];
+            tensor<fp16, [1, 64, 1536]> attn_output_15 = reshape(shape = var_1337, x = var_1318)[name = string("attn_output_15")];
+            tensor<int32, [3]> var_1342 = const()[name = string("op_1342"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_1_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310089664))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311859200))))[name = string("squeeze_1_palettized")];
+            string var_1358_pad_type_0 = const()[name = string("op_1358_pad_type_0"), val = string("valid")];
+            int32 var_1358_groups_0 = const()[name = string("op_1358_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_1358_strides_0 = const()[name = string("op_1358_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_1358_pad_0 = const()[name = string("op_1358_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_1358_dilations_0 = const()[name = string("op_1358_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_1343 = transpose(perm = var_1342, x = attn_output_15)[name = string("transpose_66")];
+            tensor<fp16, [1, 1536, 64]> var_1358 = conv(dilations = var_1358_dilations_0, groups = var_1358_groups_0, pad = var_1358_pad_0, pad_type = var_1358_pad_type_0, strides = var_1358_strides_0, weight = squeeze_1_palettized, x = var_1343)[name = string("op_1358")];
+            tensor<int32, [3]> var_1362 = const()[name = string("op_1362"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_19 = transpose(perm = var_1362, x = var_1358)[name = string("transpose_65")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_11_cast_fp16 = add(x = hidden_states_7_cast_fp16, y = attn_output_19)[name = string("hidden_states_11_cast_fp16")];
+            int32 var_1375 = const()[name = string("op_1375"), val = int32(-1)];
+            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_1377_cast_fp16 = mul(x = hidden_states_11_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_1377_cast_fp16")];
+            bool input_21_interleave_0 = const()[name = string("input_21_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_21_cast_fp16 = concat(axis = var_1375, interleave = input_21_interleave_0, values = (hidden_states_11_cast_fp16, var_1377_cast_fp16))[name = string("input_21_cast_fp16")];
+            tensor<int32, [1]> normed_13_axes_0 = const()[name = string("normed_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1372_to_fp16 = const()[name = string("op_1372_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_13_cast_fp16 = layer_norm(axes = normed_13_axes_0, epsilon = var_1372_to_fp16, x = input_21_cast_fp16)[name = string("normed_13_cast_fp16")];
+            tensor<int32, [3]> normed_15_begin_0 = const()[name = string("normed_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_15_end_0 = const()[name = string("normed_15_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_15_end_mask_0 = const()[name = string("normed_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_15_cast_fp16 = slice_by_index(begin = normed_15_begin_0, end = normed_15_end_0, end_mask = normed_15_end_mask_0, x = normed_13_cast_fp16)[name = string("normed_15_cast_fp16")];
+            tensor<fp16, [1536]> const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311908416)))];
+            tensor<fp16, [1, 64, 1536]> x_29_cast_fp16 = mul(x = normed_15_cast_fp16, y = const_52_promoted_to_fp16)[name = string("x_29_cast_fp16")];
+            tensor<int32, [3]> var_1402 = const()[name = string("op_1402"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_1403 = transpose(perm = var_1402, x = x_29_cast_fp16)[name = string("transpose_64")];
+            tensor<fp16, [1, 1536, 1, 64]> input_23 = expand_dims(axes = input_23_axes_0, x = var_1403)[name = string("input_23")];
+            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_25 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = model_model_layers_11_mlp_gate_proj_weight_palettized, x = input_23)[name = string("input_25")];
+            string b_3_pad_type_0 = const()[name = string("b_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_3_strides_0 = const()[name = string("b_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_3_pad_0 = const()[name = string("b_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_3_dilations_0 = const()[name = string("b_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_3_groups_0 = const()[name = string("b_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_3 = conv(dilations = b_3_dilations_0, groups = b_3_groups_0, pad = b_3_pad_0, pad_type = b_3_pad_type_0, strides = b_3_strides_0, weight = model_model_layers_11_mlp_up_proj_weight_palettized, x = input_23)[name = string("b_3")];
+            tensor<fp16, [1, 8960, 1, 64]> c_3 = silu(x = input_25)[name = string("c_3")];
+            tensor<fp16, [1, 8960, 1, 64]> input_27 = mul(x = c_3, y = b_3)[name = string("input_27")];
+            string e_3_pad_type_0 = const()[name = string("e_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_3_strides_0 = const()[name = string("e_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_3_pad_0 = const()[name = string("e_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_3_dilations_0 = const()[name = string("e_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_3_groups_0 = const()[name = string("e_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_3 = conv(dilations = e_3_dilations_0, groups = e_3_groups_0, pad = e_3_pad_0, pad_type = e_3_pad_type_0, strides = e_3_strides_0, weight = model_model_layers_11_mlp_down_proj_weight_palettized, x = input_27)[name = string("e_3")];
+            tensor<int32, [1]> var_1425_axes_0 = const()[name = string("op_1425_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_1425 = squeeze(axes = var_1425_axes_0, x = e_3)[name = string("op_1425")];
+            tensor<int32, [3]> var_1426 = const()[name = string("op_1426"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_1427 = transpose(perm = var_1426, x = var_1425)[name = string("transpose_63")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_13_cast_fp16 = add(x = hidden_states_11_cast_fp16, y = var_1427)[name = string("hidden_states_13_cast_fp16")];
+            int32 var_1439 = const()[name = string("op_1439"), val = int32(-1)];
+            fp16 const_53_promoted_to_fp16 = const()[name = string("const_53_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_1441_cast_fp16 = mul(x = hidden_states_13_cast_fp16, y = const_53_promoted_to_fp16)[name = string("op_1441_cast_fp16")];
+            bool input_29_interleave_0 = const()[name = string("input_29_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_29_cast_fp16 = concat(axis = var_1439, interleave = input_29_interleave_0, values = (hidden_states_13_cast_fp16, var_1441_cast_fp16))[name = string("input_29_cast_fp16")];
+            tensor<int32, [1]> normed_17_axes_0 = const()[name = string("normed_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1436_to_fp16 = const()[name = string("op_1436_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_17_cast_fp16 = layer_norm(axes = normed_17_axes_0, epsilon = var_1436_to_fp16, x = input_29_cast_fp16)[name = string("normed_17_cast_fp16")];
+            tensor<int32, [3]> normed_19_begin_0 = const()[name = string("normed_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_19_end_0 = const()[name = string("normed_19_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_19_end_mask_0 = const()[name = string("normed_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_19_cast_fp16 = slice_by_index(begin = normed_19_begin_0, end = normed_19_end_0, end_mask = normed_19_end_mask_0, x = normed_17_cast_fp16)[name = string("normed_19_cast_fp16")];
+            tensor<fp16, [1536]> const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311911552)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_15_cast_fp16 = mul(x = normed_19_cast_fp16, y = const_56_promoted_to_fp16)[name = string("hidden_states_15_cast_fp16")];
+            tensor<int32, [3]> var_1464 = const()[name = string("op_1464"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_1467_axes_0 = const()[name = string("op_1467_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_1465_cast_fp16 = transpose(perm = var_1464, x = hidden_states_15_cast_fp16)[name = string("transpose_62")];
+            tensor<fp16, [1, 1536, 1, 64]> var_1467_cast_fp16 = expand_dims(axes = var_1467_axes_0, x = var_1465_cast_fp16)[name = string("op_1467_cast_fp16")];
+            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_13 = conv(bias = model_model_layers_12_self_attn_q_proj_bias, dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = model_model_layers_12_self_attn_q_proj_weight_palettized, x = var_1467_cast_fp16)[name = string("query_states_13")];
+            string key_states_17_pad_type_0 = const()[name = string("key_states_17_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_17_strides_0 = const()[name = string("key_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_17_pad_0 = const()[name = string("key_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_17_dilations_0 = const()[name = string("key_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_17_groups_0 = const()[name = string("key_states_17_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_17 = conv(bias = model_model_layers_12_self_attn_k_proj_bias, dilations = key_states_17_dilations_0, groups = key_states_17_groups_0, pad = key_states_17_pad_0, pad_type = key_states_17_pad_type_0, strides = key_states_17_strides_0, weight = model_model_layers_12_self_attn_k_proj_weight_palettized, x = var_1467_cast_fp16)[name = string("key_states_17")];
+            string value_states_17_pad_type_0 = const()[name = string("value_states_17_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_17_strides_0 = const()[name = string("value_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_17_pad_0 = const()[name = string("value_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_17_dilations_0 = const()[name = string("value_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_17_groups_0 = const()[name = string("value_states_17_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_17 = conv(bias = model_model_layers_12_self_attn_v_proj_bias, dilations = value_states_17_dilations_0, groups = value_states_17_groups_0, pad = value_states_17_pad_0, pad_type = value_states_17_pad_type_0, strides = value_states_17_strides_0, weight = model_model_layers_12_self_attn_v_proj_weight_palettized, x = var_1467_cast_fp16)[name = string("value_states_17")];
+            tensor<int32, [4]> var_1509 = const()[name = string("op_1509"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_1510 = reshape(shape = var_1509, x = query_states_13)[name = string("op_1510")];
+            tensor<int32, [4]> var_1515 = const()[name = string("op_1515"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_1520 = const()[name = string("op_1520"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_1521 = reshape(shape = var_1520, x = key_states_17)[name = string("op_1521")];
+            tensor<int32, [4]> var_1526 = const()[name = string("op_1526"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_1532 = reshape(shape = var_1531, x = value_states_17)[name = string("op_1532")];
+            tensor<int32, [4]> var_1537 = const()[name = string("op_1537"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_9 = transpose(perm = var_1515, x = var_1510)[name = string("transpose_61")];
+            tensor<fp16, [1, 12, 64, 128]> var_1551 = mul(x = q_9, y = cos_5)[name = string("op_1551")];
+            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_9 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_9)[name = string("x1_9")];
+            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_9 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_9)[name = string("x2_9")];
+            fp16 const_60_promoted = const()[name = string("const_60_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_1572 = mul(x = x2_9, y = const_60_promoted)[name = string("op_1572")];
+            int32 var_1574 = const()[name = string("op_1574"), val = int32(-1)];
+            bool var_1575_interleave_0 = const()[name = string("op_1575_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_1575 = concat(axis = var_1574, interleave = var_1575_interleave_0, values = (var_1572, x1_9))[name = string("op_1575")];
+            tensor<fp16, [1, 12, 64, 128]> var_1576 = mul(x = var_1575, y = sin_5)[name = string("op_1576")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_15 = add(x = var_1551, y = var_1576)[name = string("query_states_15")];
+            tensor<fp16, [1, 2, 64, 128]> k_9 = transpose(perm = var_1526, x = var_1521)[name = string("transpose_60")];
+            tensor<fp16, [1, 2, 64, 128]> var_1579 = mul(x = k_9, y = cos_5)[name = string("op_1579")];
+            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_11 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_9)[name = string("x1_11")];
+            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_11 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_9)[name = string("x2_11")];
+            fp16 const_63_promoted = const()[name = string("const_63_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_1600 = mul(x = x2_11, y = const_63_promoted)[name = string("op_1600")];
+            int32 var_1602 = const()[name = string("op_1602"), val = int32(-1)];
+            bool var_1603_interleave_0 = const()[name = string("op_1603_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_1603 = concat(axis = var_1602, interleave = var_1603_interleave_0, values = (var_1600, x1_11))[name = string("op_1603")];
+            tensor<fp16, [1, 2, 64, 128]> var_1604 = mul(x = var_1603, y = sin_5)[name = string("op_1604")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_19 = add(x = var_1579, y = var_1604)[name = string("key_states_19")];
+            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([12])];
+            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([13])];
+            int32 concat_38_axis_0 = const()[name = string("concat_38_axis_0"), val = int32(0)];
+            bool concat_38_interleave_0 = const()[name = string("concat_38_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_38 = concat(axis = concat_38_axis_0, interleave = concat_38_interleave_0, values = (expand_dims_24, expand_dims_25, current_pos, expand_dims_27))[name = string("concat_38")];
+            tensor<int32, [1]> concat_39_values1_0 = const()[name = string("concat_39_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_39_values3_0 = const()[name = string("concat_39_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_39_axis_0 = const()[name = string("concat_39_axis_0"), val = int32(0)];
+            bool concat_39_interleave_0 = const()[name = string("concat_39_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_39 = concat(axis = concat_39_axis_0, interleave = concat_39_interleave_0, values = (expand_dims_28, concat_39_values1_0, var_616, concat_39_values3_0))[name = string("concat_39")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_5_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_38, begin_mask = model_model_kv_cache_0_internal_tensor_assign_5_begin_mask_0, end = concat_39, end_mask = model_model_kv_cache_0_internal_tensor_assign_5_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_5_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_5_stride_0, update = key_states_19, x = coreml_update_state_21)[name = string("model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_5_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_22_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_22 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_22")];
+            tensor<int32, [1]> expand_dims_30 = const()[name = string("expand_dims_30"), val = tensor<int32, [1]>([40])];
+            tensor<int32, [1]> expand_dims_31 = const()[name = string("expand_dims_31"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_33 = const()[name = string("expand_dims_33"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_34 = const()[name = string("expand_dims_34"), val = tensor<int32, [1]>([41])];
+            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
+            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_30, expand_dims_31, current_pos, expand_dims_33))[name = string("concat_42")];
+            tensor<int32, [1]> concat_43_values1_0 = const()[name = string("concat_43_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_43_values3_0 = const()[name = string("concat_43_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_43_axis_0 = const()[name = string("concat_43_axis_0"), val = int32(0)];
+            bool concat_43_interleave_0 = const()[name = string("concat_43_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_43 = concat(axis = concat_43_axis_0, interleave = concat_43_interleave_0, values = (expand_dims_34, concat_43_values1_0, var_616, concat_43_values3_0))[name = string("concat_43")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_6_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_19 = transpose(perm = var_1537, x = var_1532)[name = string("transpose_59")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_42, begin_mask = model_model_kv_cache_0_internal_tensor_assign_6_begin_mask_0, end = concat_43, end_mask = model_model_kv_cache_0_internal_tensor_assign_6_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_6_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_6_stride_0, update = value_states_19, x = coreml_update_state_22)[name = string("model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_6_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_23_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_23 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_23")];
+            tensor<int32, [4]> var_1675_begin_0 = const()[name = string("op_1675_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> var_1675_end_0 = const()[name = string("op_1675_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
+            tensor<bool, [4]> var_1675_end_mask_0 = const()[name = string("op_1675_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1675_cast_fp16 = slice_by_index(begin = var_1675_begin_0, end = var_1675_end_0, end_mask = var_1675_end_mask_0, x = coreml_update_state_23)[name = string("op_1675_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_5_axes_0 = const()[name = string("K_layer_cache_5_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_5_cast_fp16 = squeeze(axes = K_layer_cache_5_axes_0, x = var_1675_cast_fp16)[name = string("K_layer_cache_5_cast_fp16")];
+            tensor<int32, [4]> var_1682_begin_0 = const()[name = string("op_1682_begin_0"), val = tensor<int32, [4]>([40, 0, 0, 0])];
+            tensor<int32, [4]> var_1682_end_0 = const()[name = string("op_1682_end_0"), val = tensor<int32, [4]>([41, 2, 2048, 128])];
+            tensor<bool, [4]> var_1682_end_mask_0 = const()[name = string("op_1682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, x = coreml_update_state_23)[name = string("op_1682_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_5_axes_0 = const()[name = string("V_layer_cache_5_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_5_cast_fp16 = squeeze(axes = V_layer_cache_5_axes_0, x = var_1682_cast_fp16)[name = string("V_layer_cache_5_cast_fp16")];
+            tensor<int32, [1]> x_35_axes_0 = const()[name = string("x_35_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_35_cast_fp16 = expand_dims(axes = x_35_axes_0, x = K_layer_cache_5_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_1711 = const()[name = string("op_1711"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_37_cast_fp16 = tile(reps = var_1711, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_1723 = const()[name = string("op_1723"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_23_cast_fp16 = reshape(shape = var_1723, x = x_37_cast_fp16)[name = string("key_states_23_cast_fp16")];
+            tensor<int32, [1]> x_41_axes_0 = const()[name = string("x_41_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_41_cast_fp16 = expand_dims(axes = x_41_axes_0, x = V_layer_cache_5_cast_fp16)[name = string("x_41_cast_fp16")];
+            tensor<int32, [4]> var_1731 = const()[name = string("op_1731"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_43_cast_fp16 = tile(reps = var_1731, x = x_41_cast_fp16)[name = string("x_43_cast_fp16")];
+            bool var_1766_transpose_x_1 = const()[name = string("op_1766_transpose_x_1"), val = bool(false)];
+            bool var_1766_transpose_y_1 = const()[name = string("op_1766_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_1766_cast_fp16 = matmul(transpose_x = var_1766_transpose_x_1, transpose_y = var_1766_transpose_y_1, x = query_states_15, y = key_states_23_cast_fp16)[name = string("op_1766_cast_fp16")];
+            fp16 var_1767_to_fp16 = const()[name = string("op_1767_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_9_cast_fp16 = mul(x = var_1766_cast_fp16, y = var_1767_to_fp16)[name = string("attn_logits_9_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_11_cast_fp16 = add(x = attn_logits_9_cast_fp16, y = causal_mask)[name = string("attn_logits_11_cast_fp16")];
+            int32 var_1794 = const()[name = string("op_1794"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_1796_cast_fp16 = softmax(axis = var_1794, x = attn_logits_11_cast_fp16)[name = string("op_1796_cast_fp16")];
+            tensor<int32, [3]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_6_cast_fp16 = reshape(shape = concat_48, x = var_1796_cast_fp16)[name = string("reshape_6_cast_fp16")];
+            tensor<int32, [3]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_7_cast_fp16 = reshape(shape = concat_49, x = x_43_cast_fp16)[name = string("reshape_7_cast_fp16")];
+            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
+            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_6_cast_fp16, y = reshape_7_cast_fp16)[name = string("matmul_2_cast_fp16")];
+            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_8_cast_fp16 = reshape(shape = concat_53, x = matmul_2_cast_fp16)[name = string("reshape_8_cast_fp16")];
+            tensor<int32, [4]> var_1823_perm_0 = const()[name = string("op_1823_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1842 = const()[name = string("op_1842"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_1823 = transpose(perm = var_1823_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [1, 64, 1536]> attn_output_25 = reshape(shape = var_1842, x = var_1823)[name = string("attn_output_25")];
+            tensor<int32, [3]> var_1847 = const()[name = string("op_1847"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_2_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311914688))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313684224))))[name = string("squeeze_2_palettized")];
+            string var_1863_pad_type_0 = const()[name = string("op_1863_pad_type_0"), val = string("valid")];
+            int32 var_1863_groups_0 = const()[name = string("op_1863_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_1863_strides_0 = const()[name = string("op_1863_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_1863_pad_0 = const()[name = string("op_1863_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_1863_dilations_0 = const()[name = string("op_1863_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_1848 = transpose(perm = var_1847, x = attn_output_25)[name = string("transpose_57")];
+            tensor<fp16, [1, 1536, 64]> var_1863 = conv(dilations = var_1863_dilations_0, groups = var_1863_groups_0, pad = var_1863_pad_0, pad_type = var_1863_pad_type_0, strides = var_1863_strides_0, weight = squeeze_2_palettized, x = var_1848)[name = string("op_1863")];
+            tensor<int32, [3]> var_1867 = const()[name = string("op_1867"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_29 = transpose(perm = var_1867, x = var_1863)[name = string("transpose_56")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_17_cast_fp16 = add(x = hidden_states_13_cast_fp16, y = attn_output_29)[name = string("hidden_states_17_cast_fp16")];
+            int32 var_1880 = const()[name = string("op_1880"), val = int32(-1)];
+            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_1882_cast_fp16 = mul(x = hidden_states_17_cast_fp16, y = const_75_promoted_to_fp16)[name = string("op_1882_cast_fp16")];
+            bool input_35_interleave_0 = const()[name = string("input_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_35_cast_fp16 = concat(axis = var_1880, interleave = input_35_interleave_0, values = (hidden_states_17_cast_fp16, var_1882_cast_fp16))[name = string("input_35_cast_fp16")];
+            tensor<int32, [1]> normed_21_axes_0 = const()[name = string("normed_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1877_to_fp16 = const()[name = string("op_1877_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_21_cast_fp16 = layer_norm(axes = normed_21_axes_0, epsilon = var_1877_to_fp16, x = input_35_cast_fp16)[name = string("normed_21_cast_fp16")];
+            tensor<int32, [3]> normed_23_begin_0 = const()[name = string("normed_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_23_end_0 = const()[name = string("normed_23_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_23_end_mask_0 = const()[name = string("normed_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_23_cast_fp16 = slice_by_index(begin = normed_23_begin_0, end = normed_23_end_0, end_mask = normed_23_end_mask_0, x = normed_21_cast_fp16)[name = string("normed_23_cast_fp16")];
+            tensor<fp16, [1536]> const_78_promoted_to_fp16 = const()[name = string("const_78_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313733440)))];
+            tensor<fp16, [1, 64, 1536]> x_45_cast_fp16 = mul(x = normed_23_cast_fp16, y = const_78_promoted_to_fp16)[name = string("x_45_cast_fp16")];
+            tensor<int32, [3]> var_1907 = const()[name = string("op_1907"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_37_axes_0 = const()[name = string("input_37_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_1908 = transpose(perm = var_1907, x = x_45_cast_fp16)[name = string("transpose_55")];
+            tensor<fp16, [1, 1536, 1, 64]> input_37 = expand_dims(axes = input_37_axes_0, x = var_1908)[name = string("input_37")];
+            string input_39_pad_type_0 = const()[name = string("input_39_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_39_strides_0 = const()[name = string("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_39_pad_0 = const()[name = string("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_39_dilations_0 = const()[name = string("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_39_groups_0 = const()[name = string("input_39_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_39 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = model_model_layers_12_mlp_gate_proj_weight_palettized, x = input_37)[name = string("input_39")];
+            string b_5_pad_type_0 = const()[name = string("b_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_5_strides_0 = const()[name = string("b_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_5_pad_0 = const()[name = string("b_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_5_dilations_0 = const()[name = string("b_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_5_groups_0 = const()[name = string("b_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_5 = conv(dilations = b_5_dilations_0, groups = b_5_groups_0, pad = b_5_pad_0, pad_type = b_5_pad_type_0, strides = b_5_strides_0, weight = model_model_layers_12_mlp_up_proj_weight_palettized, x = input_37)[name = string("b_5")];
+            tensor<fp16, [1, 8960, 1, 64]> c_5 = silu(x = input_39)[name = string("c_5")];
+            tensor<fp16, [1, 8960, 1, 64]> input_41 = mul(x = c_5, y = b_5)[name = string("input_41")];
+            string e_5_pad_type_0 = const()[name = string("e_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_5_strides_0 = const()[name = string("e_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_5_pad_0 = const()[name = string("e_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_5_dilations_0 = const()[name = string("e_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_5_groups_0 = const()[name = string("e_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_5 = conv(dilations = e_5_dilations_0, groups = e_5_groups_0, pad = e_5_pad_0, pad_type = e_5_pad_type_0, strides = e_5_strides_0, weight = model_model_layers_12_mlp_down_proj_weight_palettized, x = input_41)[name = string("e_5")];
+            tensor<int32, [1]> var_1930_axes_0 = const()[name = string("op_1930_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_1930 = squeeze(axes = var_1930_axes_0, x = e_5)[name = string("op_1930")];
+            tensor<int32, [3]> var_1931 = const()[name = string("op_1931"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_1932 = transpose(perm = var_1931, x = var_1930)[name = string("transpose_54")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_19_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = var_1932)[name = string("hidden_states_19_cast_fp16")];
+            int32 var_1944 = const()[name = string("op_1944"), val = int32(-1)];
+            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_1946_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_1946_cast_fp16")];
+            bool input_43_interleave_0 = const()[name = string("input_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_43_cast_fp16 = concat(axis = var_1944, interleave = input_43_interleave_0, values = (hidden_states_19_cast_fp16, var_1946_cast_fp16))[name = string("input_43_cast_fp16")];
+            tensor<int32, [1]> normed_25_axes_0 = const()[name = string("normed_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1941_to_fp16 = const()[name = string("op_1941_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_25_cast_fp16 = layer_norm(axes = normed_25_axes_0, epsilon = var_1941_to_fp16, x = input_43_cast_fp16)[name = string("normed_25_cast_fp16")];
+            tensor<int32, [3]> normed_27_begin_0 = const()[name = string("normed_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_27_end_0 = const()[name = string("normed_27_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_27_end_mask_0 = const()[name = string("normed_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_27_cast_fp16 = slice_by_index(begin = normed_27_begin_0, end = normed_27_end_0, end_mask = normed_27_end_mask_0, x = normed_25_cast_fp16)[name = string("normed_27_cast_fp16")];
+            tensor<fp16, [1536]> const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313736576)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_21_cast_fp16 = mul(x = normed_27_cast_fp16, y = const_82_promoted_to_fp16)[name = string("hidden_states_21_cast_fp16")];
+            tensor<int32, [3]> var_1969 = const()[name = string("op_1969"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_1972_axes_0 = const()[name = string("op_1972_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_1970_cast_fp16 = transpose(perm = var_1969, x = hidden_states_21_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [1, 1536, 1, 64]> var_1972_cast_fp16 = expand_dims(axes = var_1972_axes_0, x = var_1970_cast_fp16)[name = string("op_1972_cast_fp16")];
+            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_19 = conv(bias = model_model_layers_13_self_attn_q_proj_bias, dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = model_model_layers_13_self_attn_q_proj_weight_palettized, x = var_1972_cast_fp16)[name = string("query_states_19")];
+            string key_states_25_pad_type_0 = const()[name = string("key_states_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_25_strides_0 = const()[name = string("key_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_25_pad_0 = const()[name = string("key_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_25_dilations_0 = const()[name = string("key_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_25_groups_0 = const()[name = string("key_states_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_25 = conv(bias = model_model_layers_13_self_attn_k_proj_bias, dilations = key_states_25_dilations_0, groups = key_states_25_groups_0, pad = key_states_25_pad_0, pad_type = key_states_25_pad_type_0, strides = key_states_25_strides_0, weight = model_model_layers_13_self_attn_k_proj_weight_palettized, x = var_1972_cast_fp16)[name = string("key_states_25")];
+            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_25 = conv(bias = model_model_layers_13_self_attn_v_proj_bias, dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = model_model_layers_13_self_attn_v_proj_weight_palettized, x = var_1972_cast_fp16)[name = string("value_states_25")];
+            tensor<int32, [4]> var_2014 = const()[name = string("op_2014"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_2015 = reshape(shape = var_2014, x = query_states_19)[name = string("op_2015")];
+            tensor<int32, [4]> var_2020 = const()[name = string("op_2020"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_2025 = const()[name = string("op_2025"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_2026 = reshape(shape = var_2025, x = key_states_25)[name = string("op_2026")];
+            tensor<int32, [4]> var_2031 = const()[name = string("op_2031"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_2036 = const()[name = string("op_2036"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_2037 = reshape(shape = var_2036, x = value_states_25)[name = string("op_2037")];
+            tensor<int32, [4]> var_2042 = const()[name = string("op_2042"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_13 = transpose(perm = var_2020, x = var_2015)[name = string("transpose_52")];
+            tensor<fp16, [1, 12, 64, 128]> var_2056 = mul(x = q_13, y = cos_5)[name = string("op_2056")];
+            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_13 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_13)[name = string("x1_13")];
+            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_13 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_13)[name = string("x2_13")];
+            fp16 const_86_promoted = const()[name = string("const_86_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_2077 = mul(x = x2_13, y = const_86_promoted)[name = string("op_2077")];
+            int32 var_2079 = const()[name = string("op_2079"), val = int32(-1)];
+            bool var_2080_interleave_0 = const()[name = string("op_2080_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_2080 = concat(axis = var_2079, interleave = var_2080_interleave_0, values = (var_2077, x1_13))[name = string("op_2080")];
+            tensor<fp16, [1, 12, 64, 128]> var_2081 = mul(x = var_2080, y = sin_5)[name = string("op_2081")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_21 = add(x = var_2056, y = var_2081)[name = string("query_states_21")];
+            tensor<fp16, [1, 2, 64, 128]> k_13 = transpose(perm = var_2031, x = var_2026)[name = string("transpose_51")];
+            tensor<fp16, [1, 2, 64, 128]> var_2084 = mul(x = k_13, y = cos_5)[name = string("op_2084")];
+            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_15 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_13)[name = string("x1_15")];
+            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_15 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_13)[name = string("x2_15")];
+            fp16 const_89_promoted = const()[name = string("const_89_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_2105 = mul(x = x2_15, y = const_89_promoted)[name = string("op_2105")];
+            int32 var_2107 = const()[name = string("op_2107"), val = int32(-1)];
+            bool var_2108_interleave_0 = const()[name = string("op_2108_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_2108 = concat(axis = var_2107, interleave = var_2108_interleave_0, values = (var_2105, x1_15))[name = string("op_2108")];
+            tensor<fp16, [1, 2, 64, 128]> var_2109 = mul(x = var_2108, y = sin_5)[name = string("op_2109")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_27 = add(x = var_2084, y = var_2109)[name = string("key_states_27")];
+            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([13])];
+            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([14])];
+            int32 concat_56_axis_0 = const()[name = string("concat_56_axis_0"), val = int32(0)];
+            bool concat_56_interleave_0 = const()[name = string("concat_56_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_56 = concat(axis = concat_56_axis_0, interleave = concat_56_interleave_0, values = (expand_dims_36, expand_dims_37, current_pos, expand_dims_39))[name = string("concat_56")];
+            tensor<int32, [1]> concat_57_values1_0 = const()[name = string("concat_57_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_57_values3_0 = const()[name = string("concat_57_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_57_axis_0 = const()[name = string("concat_57_axis_0"), val = int32(0)];
+            bool concat_57_interleave_0 = const()[name = string("concat_57_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_57 = concat(axis = concat_57_axis_0, interleave = concat_57_interleave_0, values = (expand_dims_40, concat_57_values1_0, var_616, concat_57_values3_0))[name = string("concat_57")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_7_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_56, begin_mask = model_model_kv_cache_0_internal_tensor_assign_7_begin_mask_0, end = concat_57, end_mask = model_model_kv_cache_0_internal_tensor_assign_7_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_7_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_7_stride_0, update = key_states_27, x = coreml_update_state_23)[name = string("model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_7_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_24_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_24 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_24")];
+            tensor<int32, [1]> expand_dims_42 = const()[name = string("expand_dims_42"), val = tensor<int32, [1]>([41])];
+            tensor<int32, [1]> expand_dims_43 = const()[name = string("expand_dims_43"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_45 = const()[name = string("expand_dims_45"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_46 = const()[name = string("expand_dims_46"), val = tensor<int32, [1]>([42])];
+            int32 concat_60_axis_0 = const()[name = string("concat_60_axis_0"), val = int32(0)];
+            bool concat_60_interleave_0 = const()[name = string("concat_60_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_60 = concat(axis = concat_60_axis_0, interleave = concat_60_interleave_0, values = (expand_dims_42, expand_dims_43, current_pos, expand_dims_45))[name = string("concat_60")];
+            tensor<int32, [1]> concat_61_values1_0 = const()[name = string("concat_61_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_61_values3_0 = const()[name = string("concat_61_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_61_axis_0 = const()[name = string("concat_61_axis_0"), val = int32(0)];
+            bool concat_61_interleave_0 = const()[name = string("concat_61_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_61 = concat(axis = concat_61_axis_0, interleave = concat_61_interleave_0, values = (expand_dims_46, concat_61_values1_0, var_616, concat_61_values3_0))[name = string("concat_61")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_8_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_27 = transpose(perm = var_2042, x = var_2037)[name = string("transpose_50")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_60, begin_mask = model_model_kv_cache_0_internal_tensor_assign_8_begin_mask_0, end = concat_61, end_mask = model_model_kv_cache_0_internal_tensor_assign_8_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_8_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_8_stride_0, update = value_states_27, x = coreml_update_state_24)[name = string("model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_8_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_25_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_25 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_25")];
+            tensor<int32, [4]> var_2180_begin_0 = const()[name = string("op_2180_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> var_2180_end_0 = const()[name = string("op_2180_end_0"), val = tensor<int32, [4]>([14, 2, 2048, 128])];
+            tensor<bool, [4]> var_2180_end_mask_0 = const()[name = string("op_2180_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2180_cast_fp16 = slice_by_index(begin = var_2180_begin_0, end = var_2180_end_0, end_mask = var_2180_end_mask_0, x = coreml_update_state_25)[name = string("op_2180_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_7_axes_0 = const()[name = string("K_layer_cache_7_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_7_cast_fp16 = squeeze(axes = K_layer_cache_7_axes_0, x = var_2180_cast_fp16)[name = string("K_layer_cache_7_cast_fp16")];
+            tensor<int32, [4]> var_2187_begin_0 = const()[name = string("op_2187_begin_0"), val = tensor<int32, [4]>([41, 0, 0, 0])];
+            tensor<int32, [4]> var_2187_end_0 = const()[name = string("op_2187_end_0"), val = tensor<int32, [4]>([42, 2, 2048, 128])];
+            tensor<bool, [4]> var_2187_end_mask_0 = const()[name = string("op_2187_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2187_cast_fp16 = slice_by_index(begin = var_2187_begin_0, end = var_2187_end_0, end_mask = var_2187_end_mask_0, x = coreml_update_state_25)[name = string("op_2187_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_7_axes_0 = const()[name = string("V_layer_cache_7_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_7_cast_fp16 = squeeze(axes = V_layer_cache_7_axes_0, x = var_2187_cast_fp16)[name = string("V_layer_cache_7_cast_fp16")];
+            tensor<int32, [1]> x_51_axes_0 = const()[name = string("x_51_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_51_cast_fp16 = expand_dims(axes = x_51_axes_0, x = K_layer_cache_7_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [4]> var_2216 = const()[name = string("op_2216"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_53_cast_fp16 = tile(reps = var_2216, x = x_51_cast_fp16)[name = string("x_53_cast_fp16")];
+            tensor<int32, [4]> var_2228 = const()[name = string("op_2228"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_31_cast_fp16 = reshape(shape = var_2228, x = x_53_cast_fp16)[name = string("key_states_31_cast_fp16")];
+            tensor<int32, [1]> x_57_axes_0 = const()[name = string("x_57_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_57_cast_fp16 = expand_dims(axes = x_57_axes_0, x = V_layer_cache_7_cast_fp16)[name = string("x_57_cast_fp16")];
+            tensor<int32, [4]> var_2236 = const()[name = string("op_2236"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_59_cast_fp16 = tile(reps = var_2236, x = x_57_cast_fp16)[name = string("x_59_cast_fp16")];
+            bool var_2271_transpose_x_1 = const()[name = string("op_2271_transpose_x_1"), val = bool(false)];
+            bool var_2271_transpose_y_1 = const()[name = string("op_2271_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_2271_cast_fp16 = matmul(transpose_x = var_2271_transpose_x_1, transpose_y = var_2271_transpose_y_1, x = query_states_21, y = key_states_31_cast_fp16)[name = string("op_2271_cast_fp16")];
+            fp16 var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_13_cast_fp16 = mul(x = var_2271_cast_fp16, y = var_2272_to_fp16)[name = string("attn_logits_13_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_15_cast_fp16 = add(x = attn_logits_13_cast_fp16, y = causal_mask)[name = string("attn_logits_15_cast_fp16")];
+            int32 var_2299 = const()[name = string("op_2299"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_2301_cast_fp16 = softmax(axis = var_2299, x = attn_logits_15_cast_fp16)[name = string("op_2301_cast_fp16")];
+            tensor<int32, [3]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_9_cast_fp16 = reshape(shape = concat_66, x = var_2301_cast_fp16)[name = string("reshape_9_cast_fp16")];
+            tensor<int32, [3]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_10_cast_fp16 = reshape(shape = concat_67, x = x_59_cast_fp16)[name = string("reshape_10_cast_fp16")];
+            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
+            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = reshape_9_cast_fp16, y = reshape_10_cast_fp16)[name = string("matmul_3_cast_fp16")];
+            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_11_cast_fp16 = reshape(shape = concat_71, x = matmul_3_cast_fp16)[name = string("reshape_11_cast_fp16")];
+            tensor<int32, [4]> var_2328_perm_0 = const()[name = string("op_2328_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2347 = const()[name = string("op_2347"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_2328 = transpose(perm = var_2328_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 64, 1536]> attn_output_35 = reshape(shape = var_2347, x = var_2328)[name = string("attn_output_35")];
+            tensor<int32, [3]> var_2352 = const()[name = string("op_2352"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_3_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313739712))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315509248))))[name = string("squeeze_3_palettized")];
+            string var_2368_pad_type_0 = const()[name = string("op_2368_pad_type_0"), val = string("valid")];
+            int32 var_2368_groups_0 = const()[name = string("op_2368_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_2368_strides_0 = const()[name = string("op_2368_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_2368_pad_0 = const()[name = string("op_2368_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_2368_dilations_0 = const()[name = string("op_2368_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_2353 = transpose(perm = var_2352, x = attn_output_35)[name = string("transpose_48")];
+            tensor<fp16, [1, 1536, 64]> var_2368 = conv(dilations = var_2368_dilations_0, groups = var_2368_groups_0, pad = var_2368_pad_0, pad_type = var_2368_pad_type_0, strides = var_2368_strides_0, weight = squeeze_3_palettized, x = var_2353)[name = string("op_2368")];
+            tensor<int32, [3]> var_2372 = const()[name = string("op_2372"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_39 = transpose(perm = var_2372, x = var_2368)[name = string("transpose_47")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_23_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = attn_output_39)[name = string("hidden_states_23_cast_fp16")];
+            int32 var_2385 = const()[name = string("op_2385"), val = int32(-1)];
+            fp16 const_101_promoted_to_fp16 = const()[name = string("const_101_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_2387_cast_fp16 = mul(x = hidden_states_23_cast_fp16, y = const_101_promoted_to_fp16)[name = string("op_2387_cast_fp16")];
+            bool input_49_interleave_0 = const()[name = string("input_49_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_49_cast_fp16 = concat(axis = var_2385, interleave = input_49_interleave_0, values = (hidden_states_23_cast_fp16, var_2387_cast_fp16))[name = string("input_49_cast_fp16")];
+            tensor<int32, [1]> normed_29_axes_0 = const()[name = string("normed_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2382_to_fp16 = const()[name = string("op_2382_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_29_cast_fp16 = layer_norm(axes = normed_29_axes_0, epsilon = var_2382_to_fp16, x = input_49_cast_fp16)[name = string("normed_29_cast_fp16")];
+            tensor<int32, [3]> normed_31_begin_0 = const()[name = string("normed_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_31_end_0 = const()[name = string("normed_31_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_31_end_mask_0 = const()[name = string("normed_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_31_cast_fp16 = slice_by_index(begin = normed_31_begin_0, end = normed_31_end_0, end_mask = normed_31_end_mask_0, x = normed_29_cast_fp16)[name = string("normed_31_cast_fp16")];
+            tensor<fp16, [1536]> const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315558464)))];
+            tensor<fp16, [1, 64, 1536]> x_61_cast_fp16 = mul(x = normed_31_cast_fp16, y = const_104_promoted_to_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [3]> var_2412 = const()[name = string("op_2412"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_2413 = transpose(perm = var_2412, x = x_61_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [1, 1536, 1, 64]> input_51 = expand_dims(axes = input_51_axes_0, x = var_2413)[name = string("input_51")];
+            string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_53 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = model_model_layers_13_mlp_gate_proj_weight_palettized, x = input_51)[name = string("input_53")];
+            string b_7_pad_type_0 = const()[name = string("b_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_7_strides_0 = const()[name = string("b_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_7_pad_0 = const()[name = string("b_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_7_dilations_0 = const()[name = string("b_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_7_groups_0 = const()[name = string("b_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_7 = conv(dilations = b_7_dilations_0, groups = b_7_groups_0, pad = b_7_pad_0, pad_type = b_7_pad_type_0, strides = b_7_strides_0, weight = model_model_layers_13_mlp_up_proj_weight_palettized, x = input_51)[name = string("b_7")];
+            tensor<fp16, [1, 8960, 1, 64]> c_7 = silu(x = input_53)[name = string("c_7")];
+            tensor<fp16, [1, 8960, 1, 64]> input_55 = mul(x = c_7, y = b_7)[name = string("input_55")];
+            string e_7_pad_type_0 = const()[name = string("e_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_7_strides_0 = const()[name = string("e_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_7_pad_0 = const()[name = string("e_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_7_dilations_0 = const()[name = string("e_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_7_groups_0 = const()[name = string("e_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_7 = conv(dilations = e_7_dilations_0, groups = e_7_groups_0, pad = e_7_pad_0, pad_type = e_7_pad_type_0, strides = e_7_strides_0, weight = model_model_layers_13_mlp_down_proj_weight_palettized, x = input_55)[name = string("e_7")];
+            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_2435 = squeeze(axes = var_2435_axes_0, x = e_7)[name = string("op_2435")];
+            tensor<int32, [3]> var_2436 = const()[name = string("op_2436"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_2437 = transpose(perm = var_2436, x = var_2435)[name = string("transpose_45")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_25_cast_fp16 = add(x = hidden_states_23_cast_fp16, y = var_2437)[name = string("hidden_states_25_cast_fp16")];
+            int32 var_2449 = const()[name = string("op_2449"), val = int32(-1)];
+            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_2451_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
+            bool input_57_interleave_0 = const()[name = string("input_57_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_57_cast_fp16 = concat(axis = var_2449, interleave = input_57_interleave_0, values = (hidden_states_25_cast_fp16, var_2451_cast_fp16))[name = string("input_57_cast_fp16")];
+            tensor<int32, [1]> normed_33_axes_0 = const()[name = string("normed_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2446_to_fp16 = const()[name = string("op_2446_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_33_cast_fp16 = layer_norm(axes = normed_33_axes_0, epsilon = var_2446_to_fp16, x = input_57_cast_fp16)[name = string("normed_33_cast_fp16")];
+            tensor<int32, [3]> normed_35_begin_0 = const()[name = string("normed_35_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_35_end_0 = const()[name = string("normed_35_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_35_end_mask_0 = const()[name = string("normed_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_35_cast_fp16 = slice_by_index(begin = normed_35_begin_0, end = normed_35_end_0, end_mask = normed_35_end_mask_0, x = normed_33_cast_fp16)[name = string("normed_35_cast_fp16")];
+            tensor<fp16, [1536]> const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315561600)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_27_cast_fp16 = mul(x = normed_35_cast_fp16, y = const_108_promoted_to_fp16)[name = string("hidden_states_27_cast_fp16")];
+            tensor<int32, [3]> var_2474 = const()[name = string("op_2474"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_2477_axes_0 = const()[name = string("op_2477_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_2475_cast_fp16 = transpose(perm = var_2474, x = hidden_states_27_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [1, 1536, 1, 64]> var_2477_cast_fp16 = expand_dims(axes = var_2477_axes_0, x = var_2475_cast_fp16)[name = string("op_2477_cast_fp16")];
+            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_25 = conv(bias = model_model_layers_14_self_attn_q_proj_bias, dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = model_model_layers_14_self_attn_q_proj_weight_palettized, x = var_2477_cast_fp16)[name = string("query_states_25")];
+            string key_states_33_pad_type_0 = const()[name = string("key_states_33_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_33_strides_0 = const()[name = string("key_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_33_pad_0 = const()[name = string("key_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_33_dilations_0 = const()[name = string("key_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_33_groups_0 = const()[name = string("key_states_33_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_33 = conv(bias = model_model_layers_14_self_attn_k_proj_bias, dilations = key_states_33_dilations_0, groups = key_states_33_groups_0, pad = key_states_33_pad_0, pad_type = key_states_33_pad_type_0, strides = key_states_33_strides_0, weight = model_model_layers_14_self_attn_k_proj_weight_palettized, x = var_2477_cast_fp16)[name = string("key_states_33")];
+            string value_states_33_pad_type_0 = const()[name = string("value_states_33_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_33_strides_0 = const()[name = string("value_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_33_pad_0 = const()[name = string("value_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_33_dilations_0 = const()[name = string("value_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_33_groups_0 = const()[name = string("value_states_33_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_33 = conv(bias = model_model_layers_14_self_attn_v_proj_bias, dilations = value_states_33_dilations_0, groups = value_states_33_groups_0, pad = value_states_33_pad_0, pad_type = value_states_33_pad_type_0, strides = value_states_33_strides_0, weight = model_model_layers_14_self_attn_v_proj_weight_palettized, x = var_2477_cast_fp16)[name = string("value_states_33")];
+            tensor<int32, [4]> var_2519 = const()[name = string("op_2519"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_2520 = reshape(shape = var_2519, x = query_states_25)[name = string("op_2520")];
+            tensor<int32, [4]> var_2525 = const()[name = string("op_2525"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_2530 = const()[name = string("op_2530"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_2531 = reshape(shape = var_2530, x = key_states_33)[name = string("op_2531")];
+            tensor<int32, [4]> var_2536 = const()[name = string("op_2536"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_2541 = const()[name = string("op_2541"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_2542 = reshape(shape = var_2541, x = value_states_33)[name = string("op_2542")];
+            tensor<int32, [4]> var_2547 = const()[name = string("op_2547"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_17 = transpose(perm = var_2525, x = var_2520)[name = string("transpose_43")];
+            tensor<fp16, [1, 12, 64, 128]> var_2561 = mul(x = q_17, y = cos_5)[name = string("op_2561")];
+            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_17 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_17)[name = string("x1_17")];
+            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_17 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_17)[name = string("x2_17")];
+            fp16 const_112_promoted = const()[name = string("const_112_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_2582 = mul(x = x2_17, y = const_112_promoted)[name = string("op_2582")];
+            int32 var_2584 = const()[name = string("op_2584"), val = int32(-1)];
+            bool var_2585_interleave_0 = const()[name = string("op_2585_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_2585 = concat(axis = var_2584, interleave = var_2585_interleave_0, values = (var_2582, x1_17))[name = string("op_2585")];
+            tensor<fp16, [1, 12, 64, 128]> var_2586 = mul(x = var_2585, y = sin_5)[name = string("op_2586")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_27 = add(x = var_2561, y = var_2586)[name = string("query_states_27")];
+            tensor<fp16, [1, 2, 64, 128]> k_17 = transpose(perm = var_2536, x = var_2531)[name = string("transpose_42")];
+            tensor<fp16, [1, 2, 64, 128]> var_2589 = mul(x = k_17, y = cos_5)[name = string("op_2589")];
+            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_19 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_17)[name = string("x1_19")];
+            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_19 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_17)[name = string("x2_19")];
+            fp16 const_115_promoted = const()[name = string("const_115_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_2610 = mul(x = x2_19, y = const_115_promoted)[name = string("op_2610")];
+            int32 var_2612 = const()[name = string("op_2612"), val = int32(-1)];
+            bool var_2613_interleave_0 = const()[name = string("op_2613_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_2613 = concat(axis = var_2612, interleave = var_2613_interleave_0, values = (var_2610, x1_19))[name = string("op_2613")];
+            tensor<fp16, [1, 2, 64, 128]> var_2614 = mul(x = var_2613, y = sin_5)[name = string("op_2614")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_35 = add(x = var_2589, y = var_2614)[name = string("key_states_35")];
+            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([14])];
+            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([15])];
+            int32 concat_74_axis_0 = const()[name = string("concat_74_axis_0"), val = int32(0)];
+            bool concat_74_interleave_0 = const()[name = string("concat_74_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_74 = concat(axis = concat_74_axis_0, interleave = concat_74_interleave_0, values = (expand_dims_48, expand_dims_49, current_pos, expand_dims_51))[name = string("concat_74")];
+            tensor<int32, [1]> concat_75_values1_0 = const()[name = string("concat_75_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_75_values3_0 = const()[name = string("concat_75_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_75_axis_0 = const()[name = string("concat_75_axis_0"), val = int32(0)];
+            bool concat_75_interleave_0 = const()[name = string("concat_75_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_75 = concat(axis = concat_75_axis_0, interleave = concat_75_interleave_0, values = (expand_dims_52, concat_75_values1_0, var_616, concat_75_values3_0))[name = string("concat_75")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_9_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_9_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_9_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_74, begin_mask = model_model_kv_cache_0_internal_tensor_assign_9_begin_mask_0, end = concat_75, end_mask = model_model_kv_cache_0_internal_tensor_assign_9_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_9_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_9_stride_0, update = key_states_35, x = coreml_update_state_25)[name = string("model_model_kv_cache_0_internal_tensor_assign_9_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_9_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_26_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_26 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_26")];
+            tensor<int32, [1]> expand_dims_54 = const()[name = string("expand_dims_54"), val = tensor<int32, [1]>([42])];
+            tensor<int32, [1]> expand_dims_55 = const()[name = string("expand_dims_55"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_57 = const()[name = string("expand_dims_57"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_58 = const()[name = string("expand_dims_58"), val = tensor<int32, [1]>([43])];
+            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
+            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_54, expand_dims_55, current_pos, expand_dims_57))[name = string("concat_78")];
+            tensor<int32, [1]> concat_79_values1_0 = const()[name = string("concat_79_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_79_values3_0 = const()[name = string("concat_79_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_79_axis_0 = const()[name = string("concat_79_axis_0"), val = int32(0)];
+            bool concat_79_interleave_0 = const()[name = string("concat_79_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_79 = concat(axis = concat_79_axis_0, interleave = concat_79_interleave_0, values = (expand_dims_58, concat_79_values1_0, var_616, concat_79_values3_0))[name = string("concat_79")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_10_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_10_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_10_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_35 = transpose(perm = var_2547, x = var_2542)[name = string("transpose_41")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_78, begin_mask = model_model_kv_cache_0_internal_tensor_assign_10_begin_mask_0, end = concat_79, end_mask = model_model_kv_cache_0_internal_tensor_assign_10_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_10_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_10_stride_0, update = value_states_35, x = coreml_update_state_26)[name = string("model_model_kv_cache_0_internal_tensor_assign_10_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_10_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_27_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_27 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_27")];
+            tensor<int32, [4]> var_2685_begin_0 = const()[name = string("op_2685_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> var_2685_end_0 = const()[name = string("op_2685_end_0"), val = tensor<int32, [4]>([15, 2, 2048, 128])];
+            tensor<bool, [4]> var_2685_end_mask_0 = const()[name = string("op_2685_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2685_cast_fp16 = slice_by_index(begin = var_2685_begin_0, end = var_2685_end_0, end_mask = var_2685_end_mask_0, x = coreml_update_state_27)[name = string("op_2685_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_9_axes_0 = const()[name = string("K_layer_cache_9_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_9_cast_fp16 = squeeze(axes = K_layer_cache_9_axes_0, x = var_2685_cast_fp16)[name = string("K_layer_cache_9_cast_fp16")];
+            tensor<int32, [4]> var_2692_begin_0 = const()[name = string("op_2692_begin_0"), val = tensor<int32, [4]>([42, 0, 0, 0])];
+            tensor<int32, [4]> var_2692_end_0 = const()[name = string("op_2692_end_0"), val = tensor<int32, [4]>([43, 2, 2048, 128])];
+            tensor<bool, [4]> var_2692_end_mask_0 = const()[name = string("op_2692_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_2692_cast_fp16 = slice_by_index(begin = var_2692_begin_0, end = var_2692_end_0, end_mask = var_2692_end_mask_0, x = coreml_update_state_27)[name = string("op_2692_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_9_axes_0 = const()[name = string("V_layer_cache_9_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_9_cast_fp16 = squeeze(axes = V_layer_cache_9_axes_0, x = var_2692_cast_fp16)[name = string("V_layer_cache_9_cast_fp16")];
+            tensor<int32, [1]> x_67_axes_0 = const()[name = string("x_67_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_67_cast_fp16 = expand_dims(axes = x_67_axes_0, x = K_layer_cache_9_cast_fp16)[name = string("x_67_cast_fp16")];
+            tensor<int32, [4]> var_2721 = const()[name = string("op_2721"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_69_cast_fp16 = tile(reps = var_2721, x = x_67_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [4]> var_2733 = const()[name = string("op_2733"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_39_cast_fp16 = reshape(shape = var_2733, x = x_69_cast_fp16)[name = string("key_states_39_cast_fp16")];
+            tensor<int32, [1]> x_73_axes_0 = const()[name = string("x_73_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_73_cast_fp16 = expand_dims(axes = x_73_axes_0, x = V_layer_cache_9_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [4]> var_2741 = const()[name = string("op_2741"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_75_cast_fp16 = tile(reps = var_2741, x = x_73_cast_fp16)[name = string("x_75_cast_fp16")];
+            bool var_2776_transpose_x_1 = const()[name = string("op_2776_transpose_x_1"), val = bool(false)];
+            bool var_2776_transpose_y_1 = const()[name = string("op_2776_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_2776_cast_fp16 = matmul(transpose_x = var_2776_transpose_x_1, transpose_y = var_2776_transpose_y_1, x = query_states_27, y = key_states_39_cast_fp16)[name = string("op_2776_cast_fp16")];
+            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_17_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("attn_logits_17_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_19_cast_fp16 = add(x = attn_logits_17_cast_fp16, y = causal_mask)[name = string("attn_logits_19_cast_fp16")];
+            int32 var_2804 = const()[name = string("op_2804"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_2806_cast_fp16 = softmax(axis = var_2804, x = attn_logits_19_cast_fp16)[name = string("op_2806_cast_fp16")];
+            tensor<int32, [3]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_12_cast_fp16 = reshape(shape = concat_84, x = var_2806_cast_fp16)[name = string("reshape_12_cast_fp16")];
+            tensor<int32, [3]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_13_cast_fp16 = reshape(shape = concat_85, x = x_75_cast_fp16)[name = string("reshape_13_cast_fp16")];
+            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
+            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = reshape_12_cast_fp16, y = reshape_13_cast_fp16)[name = string("matmul_4_cast_fp16")];
+            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_14_cast_fp16 = reshape(shape = concat_89, x = matmul_4_cast_fp16)[name = string("reshape_14_cast_fp16")];
+            tensor<int32, [4]> var_2833_perm_0 = const()[name = string("op_2833_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2852 = const()[name = string("op_2852"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_2833 = transpose(perm = var_2833_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [1, 64, 1536]> attn_output_45 = reshape(shape = var_2852, x = var_2833)[name = string("attn_output_45")];
+            tensor<int32, [3]> var_2857 = const()[name = string("op_2857"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_4_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315564736))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317334272))))[name = string("squeeze_4_palettized")];
+            string var_2873_pad_type_0 = const()[name = string("op_2873_pad_type_0"), val = string("valid")];
+            int32 var_2873_groups_0 = const()[name = string("op_2873_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_2873_strides_0 = const()[name = string("op_2873_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_2873_pad_0 = const()[name = string("op_2873_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_2873_dilations_0 = const()[name = string("op_2873_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_2858 = transpose(perm = var_2857, x = attn_output_45)[name = string("transpose_39")];
+            tensor<fp16, [1, 1536, 64]> var_2873 = conv(dilations = var_2873_dilations_0, groups = var_2873_groups_0, pad = var_2873_pad_0, pad_type = var_2873_pad_type_0, strides = var_2873_strides_0, weight = squeeze_4_palettized, x = var_2858)[name = string("op_2873")];
+            tensor<int32, [3]> var_2877 = const()[name = string("op_2877"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_49 = transpose(perm = var_2877, x = var_2873)[name = string("transpose_38")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = attn_output_49)[name = string("hidden_states_29_cast_fp16")];
+            int32 var_2890 = const()[name = string("op_2890"), val = int32(-1)];
+            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_2892_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_127_promoted_to_fp16)[name = string("op_2892_cast_fp16")];
+            bool input_63_interleave_0 = const()[name = string("input_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_63_cast_fp16 = concat(axis = var_2890, interleave = input_63_interleave_0, values = (hidden_states_29_cast_fp16, var_2892_cast_fp16))[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> normed_37_axes_0 = const()[name = string("normed_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2887_to_fp16 = const()[name = string("op_2887_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_37_cast_fp16 = layer_norm(axes = normed_37_axes_0, epsilon = var_2887_to_fp16, x = input_63_cast_fp16)[name = string("normed_37_cast_fp16")];
+            tensor<int32, [3]> normed_39_begin_0 = const()[name = string("normed_39_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_39_end_0 = const()[name = string("normed_39_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_39_end_mask_0 = const()[name = string("normed_39_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_39_cast_fp16 = slice_by_index(begin = normed_39_begin_0, end = normed_39_end_0, end_mask = normed_39_end_mask_0, x = normed_37_cast_fp16)[name = string("normed_39_cast_fp16")];
+            tensor<fp16, [1536]> const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317383488)))];
+            tensor<fp16, [1, 64, 1536]> x_77_cast_fp16 = mul(x = normed_39_cast_fp16, y = const_130_promoted_to_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> var_2917 = const()[name = string("op_2917"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_2918 = transpose(perm = var_2917, x = x_77_cast_fp16)[name = string("transpose_37")];
+            tensor<fp16, [1, 1536, 1, 64]> input_65 = expand_dims(axes = input_65_axes_0, x = var_2918)[name = string("input_65")];
+            string input_67_pad_type_0 = const()[name = string("input_67_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_67_strides_0 = const()[name = string("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_67_dilations_0 = const()[name = string("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_67_groups_0 = const()[name = string("input_67_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_67 = conv(dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = model_model_layers_14_mlp_gate_proj_weight_palettized, x = input_65)[name = string("input_67")];
+            string b_9_pad_type_0 = const()[name = string("b_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_9_strides_0 = const()[name = string("b_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_9_pad_0 = const()[name = string("b_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_9_dilations_0 = const()[name = string("b_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_9_groups_0 = const()[name = string("b_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_9 = conv(dilations = b_9_dilations_0, groups = b_9_groups_0, pad = b_9_pad_0, pad_type = b_9_pad_type_0, strides = b_9_strides_0, weight = model_model_layers_14_mlp_up_proj_weight_palettized, x = input_65)[name = string("b_9")];
+            tensor<fp16, [1, 8960, 1, 64]> c_9 = silu(x = input_67)[name = string("c_9")];
+            tensor<fp16, [1, 8960, 1, 64]> input_69 = mul(x = c_9, y = b_9)[name = string("input_69")];
+            string e_9_pad_type_0 = const()[name = string("e_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_9_strides_0 = const()[name = string("e_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_9_pad_0 = const()[name = string("e_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_9_dilations_0 = const()[name = string("e_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_9_groups_0 = const()[name = string("e_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_9 = conv(dilations = e_9_dilations_0, groups = e_9_groups_0, pad = e_9_pad_0, pad_type = e_9_pad_type_0, strides = e_9_strides_0, weight = model_model_layers_14_mlp_down_proj_weight_palettized, x = input_69)[name = string("e_9")];
+            tensor<int32, [1]> var_2940_axes_0 = const()[name = string("op_2940_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_2940 = squeeze(axes = var_2940_axes_0, x = e_9)[name = string("op_2940")];
+            tensor<int32, [3]> var_2941 = const()[name = string("op_2941"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_2942 = transpose(perm = var_2941, x = var_2940)[name = string("transpose_36")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_31_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = var_2942)[name = string("hidden_states_31_cast_fp16")];
+            int32 var_2954 = const()[name = string("op_2954"), val = int32(-1)];
+            fp16 const_131_promoted_to_fp16 = const()[name = string("const_131_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_2956_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = const_131_promoted_to_fp16)[name = string("op_2956_cast_fp16")];
+            bool input_71_interleave_0 = const()[name = string("input_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_71_cast_fp16 = concat(axis = var_2954, interleave = input_71_interleave_0, values = (hidden_states_31_cast_fp16, var_2956_cast_fp16))[name = string("input_71_cast_fp16")];
+            tensor<int32, [1]> normed_41_axes_0 = const()[name = string("normed_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2951_to_fp16 = const()[name = string("op_2951_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_41_cast_fp16 = layer_norm(axes = normed_41_axes_0, epsilon = var_2951_to_fp16, x = input_71_cast_fp16)[name = string("normed_41_cast_fp16")];
+            tensor<int32, [3]> normed_43_begin_0 = const()[name = string("normed_43_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_43_end_0 = const()[name = string("normed_43_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_43_end_mask_0 = const()[name = string("normed_43_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_43_cast_fp16 = slice_by_index(begin = normed_43_begin_0, end = normed_43_end_0, end_mask = normed_43_end_mask_0, x = normed_41_cast_fp16)[name = string("normed_43_cast_fp16")];
+            tensor<fp16, [1536]> const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317386624)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_33_cast_fp16 = mul(x = normed_43_cast_fp16, y = const_134_promoted_to_fp16)[name = string("hidden_states_33_cast_fp16")];
+            tensor<int32, [3]> var_2979 = const()[name = string("op_2979"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_2982_axes_0 = const()[name = string("op_2982_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_2980_cast_fp16 = transpose(perm = var_2979, x = hidden_states_33_cast_fp16)[name = string("transpose_35")];
+            tensor<fp16, [1, 1536, 1, 64]> var_2982_cast_fp16 = expand_dims(axes = var_2982_axes_0, x = var_2980_cast_fp16)[name = string("op_2982_cast_fp16")];
+            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_31 = conv(bias = model_model_layers_15_self_attn_q_proj_bias, dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = model_model_layers_15_self_attn_q_proj_weight_palettized, x = var_2982_cast_fp16)[name = string("query_states_31")];
+            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_41 = conv(bias = model_model_layers_15_self_attn_k_proj_bias, dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = model_model_layers_15_self_attn_k_proj_weight_palettized, x = var_2982_cast_fp16)[name = string("key_states_41")];
+            string value_states_41_pad_type_0 = const()[name = string("value_states_41_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_41_strides_0 = const()[name = string("value_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_41_pad_0 = const()[name = string("value_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_41_dilations_0 = const()[name = string("value_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_41_groups_0 = const()[name = string("value_states_41_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_41 = conv(bias = model_model_layers_15_self_attn_v_proj_bias, dilations = value_states_41_dilations_0, groups = value_states_41_groups_0, pad = value_states_41_pad_0, pad_type = value_states_41_pad_type_0, strides = value_states_41_strides_0, weight = model_model_layers_15_self_attn_v_proj_weight_palettized, x = var_2982_cast_fp16)[name = string("value_states_41")];
+            tensor<int32, [4]> var_3024 = const()[name = string("op_3024"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_3025 = reshape(shape = var_3024, x = query_states_31)[name = string("op_3025")];
+            tensor<int32, [4]> var_3030 = const()[name = string("op_3030"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_3035 = const()[name = string("op_3035"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_3036 = reshape(shape = var_3035, x = key_states_41)[name = string("op_3036")];
+            tensor<int32, [4]> var_3041 = const()[name = string("op_3041"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_3046 = const()[name = string("op_3046"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_3047 = reshape(shape = var_3046, x = value_states_41)[name = string("op_3047")];
+            tensor<int32, [4]> var_3052 = const()[name = string("op_3052"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_21 = transpose(perm = var_3030, x = var_3025)[name = string("transpose_34")];
+            tensor<fp16, [1, 12, 64, 128]> var_3066 = mul(x = q_21, y = cos_5)[name = string("op_3066")];
+            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_21 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_21)[name = string("x1_21")];
+            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_21 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_21)[name = string("x2_21")];
+            fp16 const_138_promoted = const()[name = string("const_138_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_3087 = mul(x = x2_21, y = const_138_promoted)[name = string("op_3087")];
+            int32 var_3089 = const()[name = string("op_3089"), val = int32(-1)];
+            bool var_3090_interleave_0 = const()[name = string("op_3090_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_3090 = concat(axis = var_3089, interleave = var_3090_interleave_0, values = (var_3087, x1_21))[name = string("op_3090")];
+            tensor<fp16, [1, 12, 64, 128]> var_3091 = mul(x = var_3090, y = sin_5)[name = string("op_3091")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_33 = add(x = var_3066, y = var_3091)[name = string("query_states_33")];
+            tensor<fp16, [1, 2, 64, 128]> k_21 = transpose(perm = var_3041, x = var_3036)[name = string("transpose_33")];
+            tensor<fp16, [1, 2, 64, 128]> var_3094 = mul(x = k_21, y = cos_5)[name = string("op_3094")];
+            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_23 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_21)[name = string("x1_23")];
+            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_23 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_21)[name = string("x2_23")];
+            fp16 const_141_promoted = const()[name = string("const_141_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_3115 = mul(x = x2_23, y = const_141_promoted)[name = string("op_3115")];
+            int32 var_3117 = const()[name = string("op_3117"), val = int32(-1)];
+            bool var_3118_interleave_0 = const()[name = string("op_3118_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_3118 = concat(axis = var_3117, interleave = var_3118_interleave_0, values = (var_3115, x1_23))[name = string("op_3118")];
+            tensor<fp16, [1, 2, 64, 128]> var_3119 = mul(x = var_3118, y = sin_5)[name = string("op_3119")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_43 = add(x = var_3094, y = var_3119)[name = string("key_states_43")];
+            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([15])];
+            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([16])];
+            int32 concat_92_axis_0 = const()[name = string("concat_92_axis_0"), val = int32(0)];
+            bool concat_92_interleave_0 = const()[name = string("concat_92_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_92 = concat(axis = concat_92_axis_0, interleave = concat_92_interleave_0, values = (expand_dims_60, expand_dims_61, current_pos, expand_dims_63))[name = string("concat_92")];
+            tensor<int32, [1]> concat_93_values1_0 = const()[name = string("concat_93_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_93_values3_0 = const()[name = string("concat_93_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_93_axis_0 = const()[name = string("concat_93_axis_0"), val = int32(0)];
+            bool concat_93_interleave_0 = const()[name = string("concat_93_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_93 = concat(axis = concat_93_axis_0, interleave = concat_93_interleave_0, values = (expand_dims_64, concat_93_values1_0, var_616, concat_93_values3_0))[name = string("concat_93")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_11_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_11_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_11_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_92, begin_mask = model_model_kv_cache_0_internal_tensor_assign_11_begin_mask_0, end = concat_93, end_mask = model_model_kv_cache_0_internal_tensor_assign_11_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_11_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_11_stride_0, update = key_states_43, x = coreml_update_state_27)[name = string("model_model_kv_cache_0_internal_tensor_assign_11_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_11_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_28_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_28 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_28")];
+            tensor<int32, [1]> expand_dims_66 = const()[name = string("expand_dims_66"), val = tensor<int32, [1]>([43])];
+            tensor<int32, [1]> expand_dims_67 = const()[name = string("expand_dims_67"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_69 = const()[name = string("expand_dims_69"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_70 = const()[name = string("expand_dims_70"), val = tensor<int32, [1]>([44])];
+            int32 concat_96_axis_0 = const()[name = string("concat_96_axis_0"), val = int32(0)];
+            bool concat_96_interleave_0 = const()[name = string("concat_96_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_96 = concat(axis = concat_96_axis_0, interleave = concat_96_interleave_0, values = (expand_dims_66, expand_dims_67, current_pos, expand_dims_69))[name = string("concat_96")];
+            tensor<int32, [1]> concat_97_values1_0 = const()[name = string("concat_97_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_97_values3_0 = const()[name = string("concat_97_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_97_axis_0 = const()[name = string("concat_97_axis_0"), val = int32(0)];
+            bool concat_97_interleave_0 = const()[name = string("concat_97_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_97 = concat(axis = concat_97_axis_0, interleave = concat_97_interleave_0, values = (expand_dims_70, concat_97_values1_0, var_616, concat_97_values3_0))[name = string("concat_97")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_12_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_12_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_12_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_43 = transpose(perm = var_3052, x = var_3047)[name = string("transpose_32")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_96, begin_mask = model_model_kv_cache_0_internal_tensor_assign_12_begin_mask_0, end = concat_97, end_mask = model_model_kv_cache_0_internal_tensor_assign_12_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_12_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_12_stride_0, update = value_states_43, x = coreml_update_state_28)[name = string("model_model_kv_cache_0_internal_tensor_assign_12_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_12_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_29_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_29 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_29")];
+            tensor<int32, [4]> var_3190_begin_0 = const()[name = string("op_3190_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> var_3190_end_0 = const()[name = string("op_3190_end_0"), val = tensor<int32, [4]>([16, 2, 2048, 128])];
+            tensor<bool, [4]> var_3190_end_mask_0 = const()[name = string("op_3190_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3190_cast_fp16 = slice_by_index(begin = var_3190_begin_0, end = var_3190_end_0, end_mask = var_3190_end_mask_0, x = coreml_update_state_29)[name = string("op_3190_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_11_axes_0 = const()[name = string("K_layer_cache_11_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_11_cast_fp16 = squeeze(axes = K_layer_cache_11_axes_0, x = var_3190_cast_fp16)[name = string("K_layer_cache_11_cast_fp16")];
+            tensor<int32, [4]> var_3197_begin_0 = const()[name = string("op_3197_begin_0"), val = tensor<int32, [4]>([43, 0, 0, 0])];
+            tensor<int32, [4]> var_3197_end_0 = const()[name = string("op_3197_end_0"), val = tensor<int32, [4]>([44, 2, 2048, 128])];
+            tensor<bool, [4]> var_3197_end_mask_0 = const()[name = string("op_3197_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3197_cast_fp16 = slice_by_index(begin = var_3197_begin_0, end = var_3197_end_0, end_mask = var_3197_end_mask_0, x = coreml_update_state_29)[name = string("op_3197_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_11_axes_0 = const()[name = string("V_layer_cache_11_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_11_cast_fp16 = squeeze(axes = V_layer_cache_11_axes_0, x = var_3197_cast_fp16)[name = string("V_layer_cache_11_cast_fp16")];
+            tensor<int32, [1]> x_83_axes_0 = const()[name = string("x_83_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_83_cast_fp16 = expand_dims(axes = x_83_axes_0, x = K_layer_cache_11_cast_fp16)[name = string("x_83_cast_fp16")];
+            tensor<int32, [4]> var_3226 = const()[name = string("op_3226"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_85_cast_fp16 = tile(reps = var_3226, x = x_83_cast_fp16)[name = string("x_85_cast_fp16")];
+            tensor<int32, [4]> var_3238 = const()[name = string("op_3238"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_47_cast_fp16 = reshape(shape = var_3238, x = x_85_cast_fp16)[name = string("key_states_47_cast_fp16")];
+            tensor<int32, [1]> x_89_axes_0 = const()[name = string("x_89_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_89_cast_fp16 = expand_dims(axes = x_89_axes_0, x = V_layer_cache_11_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_3246 = const()[name = string("op_3246"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_91_cast_fp16 = tile(reps = var_3246, x = x_89_cast_fp16)[name = string("x_91_cast_fp16")];
+            bool var_3281_transpose_x_1 = const()[name = string("op_3281_transpose_x_1"), val = bool(false)];
+            bool var_3281_transpose_y_1 = const()[name = string("op_3281_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_3281_cast_fp16 = matmul(transpose_x = var_3281_transpose_x_1, transpose_y = var_3281_transpose_y_1, x = query_states_33, y = key_states_47_cast_fp16)[name = string("op_3281_cast_fp16")];
+            fp16 var_3282_to_fp16 = const()[name = string("op_3282_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_21_cast_fp16 = mul(x = var_3281_cast_fp16, y = var_3282_to_fp16)[name = string("attn_logits_21_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_23_cast_fp16 = add(x = attn_logits_21_cast_fp16, y = causal_mask)[name = string("attn_logits_23_cast_fp16")];
+            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_3311_cast_fp16 = softmax(axis = var_3309, x = attn_logits_23_cast_fp16)[name = string("op_3311_cast_fp16")];
+            tensor<int32, [3]> concat_102 = const()[name = string("concat_102"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_15_cast_fp16 = reshape(shape = concat_102, x = var_3311_cast_fp16)[name = string("reshape_15_cast_fp16")];
+            tensor<int32, [3]> concat_103 = const()[name = string("concat_103"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_16_cast_fp16 = reshape(shape = concat_103, x = x_91_cast_fp16)[name = string("reshape_16_cast_fp16")];
+            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
+            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = reshape_15_cast_fp16, y = reshape_16_cast_fp16)[name = string("matmul_5_cast_fp16")];
+            tensor<int32, [4]> concat_107 = const()[name = string("concat_107"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_17_cast_fp16 = reshape(shape = concat_107, x = matmul_5_cast_fp16)[name = string("reshape_17_cast_fp16")];
+            tensor<int32, [4]> var_3338_perm_0 = const()[name = string("op_3338_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3357 = const()[name = string("op_3357"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_3338 = transpose(perm = var_3338_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_31")];
+            tensor<fp16, [1, 64, 1536]> attn_output_55 = reshape(shape = var_3357, x = var_3338)[name = string("attn_output_55")];
+            tensor<int32, [3]> var_3362 = const()[name = string("op_3362"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_5_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317389760))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319159296))))[name = string("squeeze_5_palettized")];
+            string var_3378_pad_type_0 = const()[name = string("op_3378_pad_type_0"), val = string("valid")];
+            int32 var_3378_groups_0 = const()[name = string("op_3378_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_3378_strides_0 = const()[name = string("op_3378_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_3378_pad_0 = const()[name = string("op_3378_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_3378_dilations_0 = const()[name = string("op_3378_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_3363 = transpose(perm = var_3362, x = attn_output_55)[name = string("transpose_30")];
+            tensor<fp16, [1, 1536, 64]> var_3378 = conv(dilations = var_3378_dilations_0, groups = var_3378_groups_0, pad = var_3378_pad_0, pad_type = var_3378_pad_type_0, strides = var_3378_strides_0, weight = squeeze_5_palettized, x = var_3363)[name = string("op_3378")];
+            tensor<int32, [3]> var_3382 = const()[name = string("op_3382"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_59 = transpose(perm = var_3382, x = var_3378)[name = string("transpose_29")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_35_cast_fp16 = add(x = hidden_states_31_cast_fp16, y = attn_output_59)[name = string("hidden_states_35_cast_fp16")];
+            int32 var_3395 = const()[name = string("op_3395"), val = int32(-1)];
+            fp16 const_153_promoted_to_fp16 = const()[name = string("const_153_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_3397_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_153_promoted_to_fp16)[name = string("op_3397_cast_fp16")];
+            bool input_77_interleave_0 = const()[name = string("input_77_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_77_cast_fp16 = concat(axis = var_3395, interleave = input_77_interleave_0, values = (hidden_states_35_cast_fp16, var_3397_cast_fp16))[name = string("input_77_cast_fp16")];
+            tensor<int32, [1]> normed_45_axes_0 = const()[name = string("normed_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3392_to_fp16 = const()[name = string("op_3392_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_45_cast_fp16 = layer_norm(axes = normed_45_axes_0, epsilon = var_3392_to_fp16, x = input_77_cast_fp16)[name = string("normed_45_cast_fp16")];
+            tensor<int32, [3]> normed_47_begin_0 = const()[name = string("normed_47_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_47_end_0 = const()[name = string("normed_47_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_47_end_mask_0 = const()[name = string("normed_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_47_cast_fp16 = slice_by_index(begin = normed_47_begin_0, end = normed_47_end_0, end_mask = normed_47_end_mask_0, x = normed_45_cast_fp16)[name = string("normed_47_cast_fp16")];
+            tensor<fp16, [1536]> const_156_promoted_to_fp16 = const()[name = string("const_156_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319208512)))];
+            tensor<fp16, [1, 64, 1536]> x_93_cast_fp16 = mul(x = normed_47_cast_fp16, y = const_156_promoted_to_fp16)[name = string("x_93_cast_fp16")];
+            tensor<int32, [3]> var_3422 = const()[name = string("op_3422"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_3423 = transpose(perm = var_3422, x = x_93_cast_fp16)[name = string("transpose_28")];
+            tensor<fp16, [1, 1536, 1, 64]> input_79 = expand_dims(axes = input_79_axes_0, x = var_3423)[name = string("input_79")];
+            string input_81_pad_type_0 = const()[name = string("input_81_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_81_strides_0 = const()[name = string("input_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_81_pad_0 = const()[name = string("input_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_81_dilations_0 = const()[name = string("input_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_81_groups_0 = const()[name = string("input_81_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_81 = conv(dilations = input_81_dilations_0, groups = input_81_groups_0, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = input_81_strides_0, weight = model_model_layers_15_mlp_gate_proj_weight_palettized, x = input_79)[name = string("input_81")];
+            string b_11_pad_type_0 = const()[name = string("b_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_11_strides_0 = const()[name = string("b_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_11_pad_0 = const()[name = string("b_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_11_dilations_0 = const()[name = string("b_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_11_groups_0 = const()[name = string("b_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_11 = conv(dilations = b_11_dilations_0, groups = b_11_groups_0, pad = b_11_pad_0, pad_type = b_11_pad_type_0, strides = b_11_strides_0, weight = model_model_layers_15_mlp_up_proj_weight_palettized, x = input_79)[name = string("b_11")];
+            tensor<fp16, [1, 8960, 1, 64]> c_11 = silu(x = input_81)[name = string("c_11")];
+            tensor<fp16, [1, 8960, 1, 64]> input_83 = mul(x = c_11, y = b_11)[name = string("input_83")];
+            string e_11_pad_type_0 = const()[name = string("e_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_11_strides_0 = const()[name = string("e_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_11_pad_0 = const()[name = string("e_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_11_dilations_0 = const()[name = string("e_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_11_groups_0 = const()[name = string("e_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_11 = conv(dilations = e_11_dilations_0, groups = e_11_groups_0, pad = e_11_pad_0, pad_type = e_11_pad_type_0, strides = e_11_strides_0, weight = model_model_layers_15_mlp_down_proj_weight_palettized, x = input_83)[name = string("e_11")];
+            tensor<int32, [1]> var_3445_axes_0 = const()[name = string("op_3445_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_3445 = squeeze(axes = var_3445_axes_0, x = e_11)[name = string("op_3445")];
+            tensor<int32, [3]> var_3446 = const()[name = string("op_3446"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_3447 = transpose(perm = var_3446, x = var_3445)[name = string("transpose_27")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_37_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = var_3447)[name = string("hidden_states_37_cast_fp16")];
+            int32 var_3459 = const()[name = string("op_3459"), val = int32(-1)];
+            fp16 const_157_promoted_to_fp16 = const()[name = string("const_157_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_3461_cast_fp16 = mul(x = hidden_states_37_cast_fp16, y = const_157_promoted_to_fp16)[name = string("op_3461_cast_fp16")];
+            bool input_85_interleave_0 = const()[name = string("input_85_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_85_cast_fp16 = concat(axis = var_3459, interleave = input_85_interleave_0, values = (hidden_states_37_cast_fp16, var_3461_cast_fp16))[name = string("input_85_cast_fp16")];
+            tensor<int32, [1]> normed_49_axes_0 = const()[name = string("normed_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3456_to_fp16 = const()[name = string("op_3456_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_49_cast_fp16 = layer_norm(axes = normed_49_axes_0, epsilon = var_3456_to_fp16, x = input_85_cast_fp16)[name = string("normed_49_cast_fp16")];
+            tensor<int32, [3]> normed_51_begin_0 = const()[name = string("normed_51_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_51_end_0 = const()[name = string("normed_51_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_51_end_mask_0 = const()[name = string("normed_51_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_51_cast_fp16 = slice_by_index(begin = normed_51_begin_0, end = normed_51_end_0, end_mask = normed_51_end_mask_0, x = normed_49_cast_fp16)[name = string("normed_51_cast_fp16")];
+            tensor<fp16, [1536]> const_160_promoted_to_fp16 = const()[name = string("const_160_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319211648)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_39_cast_fp16 = mul(x = normed_51_cast_fp16, y = const_160_promoted_to_fp16)[name = string("hidden_states_39_cast_fp16")];
+            tensor<int32, [3]> var_3484 = const()[name = string("op_3484"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_3487_axes_0 = const()[name = string("op_3487_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_3485_cast_fp16 = transpose(perm = var_3484, x = hidden_states_39_cast_fp16)[name = string("transpose_26")];
+            tensor<fp16, [1, 1536, 1, 64]> var_3487_cast_fp16 = expand_dims(axes = var_3487_axes_0, x = var_3485_cast_fp16)[name = string("op_3487_cast_fp16")];
+            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_37 = conv(bias = model_model_layers_16_self_attn_q_proj_bias, dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = model_model_layers_16_self_attn_q_proj_weight_palettized, x = var_3487_cast_fp16)[name = string("query_states_37")];
+            string key_states_49_pad_type_0 = const()[name = string("key_states_49_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_49_strides_0 = const()[name = string("key_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_49_pad_0 = const()[name = string("key_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_49_dilations_0 = const()[name = string("key_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_49_groups_0 = const()[name = string("key_states_49_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_49 = conv(bias = model_model_layers_16_self_attn_k_proj_bias, dilations = key_states_49_dilations_0, groups = key_states_49_groups_0, pad = key_states_49_pad_0, pad_type = key_states_49_pad_type_0, strides = key_states_49_strides_0, weight = model_model_layers_16_self_attn_k_proj_weight_palettized, x = var_3487_cast_fp16)[name = string("key_states_49")];
+            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_49 = conv(bias = model_model_layers_16_self_attn_v_proj_bias, dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = model_model_layers_16_self_attn_v_proj_weight_palettized, x = var_3487_cast_fp16)[name = string("value_states_49")];
+            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_3530 = reshape(shape = var_3529, x = query_states_37)[name = string("op_3530")];
+            tensor<int32, [4]> var_3535 = const()[name = string("op_3535"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_3540 = const()[name = string("op_3540"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_3541 = reshape(shape = var_3540, x = key_states_49)[name = string("op_3541")];
+            tensor<int32, [4]> var_3546 = const()[name = string("op_3546"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_3551 = const()[name = string("op_3551"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_3552 = reshape(shape = var_3551, x = value_states_49)[name = string("op_3552")];
+            tensor<int32, [4]> var_3557 = const()[name = string("op_3557"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_25 = transpose(perm = var_3535, x = var_3530)[name = string("transpose_25")];
+            tensor<fp16, [1, 12, 64, 128]> var_3571 = mul(x = q_25, y = cos_5)[name = string("op_3571")];
+            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_25 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_25)[name = string("x1_25")];
+            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_25 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_25)[name = string("x2_25")];
+            fp16 const_164_promoted = const()[name = string("const_164_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_3592 = mul(x = x2_25, y = const_164_promoted)[name = string("op_3592")];
+            int32 var_3594 = const()[name = string("op_3594"), val = int32(-1)];
+            bool var_3595_interleave_0 = const()[name = string("op_3595_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_3595 = concat(axis = var_3594, interleave = var_3595_interleave_0, values = (var_3592, x1_25))[name = string("op_3595")];
+            tensor<fp16, [1, 12, 64, 128]> var_3596 = mul(x = var_3595, y = sin_5)[name = string("op_3596")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_39 = add(x = var_3571, y = var_3596)[name = string("query_states_39")];
+            tensor<fp16, [1, 2, 64, 128]> k_25 = transpose(perm = var_3546, x = var_3541)[name = string("transpose_24")];
+            tensor<fp16, [1, 2, 64, 128]> var_3599 = mul(x = k_25, y = cos_5)[name = string("op_3599")];
+            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_27 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_25)[name = string("x1_27")];
+            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_27 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_25)[name = string("x2_27")];
+            fp16 const_167_promoted = const()[name = string("const_167_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_3620 = mul(x = x2_27, y = const_167_promoted)[name = string("op_3620")];
+            int32 var_3622 = const()[name = string("op_3622"), val = int32(-1)];
+            bool var_3623_interleave_0 = const()[name = string("op_3623_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_3623 = concat(axis = var_3622, interleave = var_3623_interleave_0, values = (var_3620, x1_27))[name = string("op_3623")];
+            tensor<fp16, [1, 2, 64, 128]> var_3624 = mul(x = var_3623, y = sin_5)[name = string("op_3624")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_51 = add(x = var_3599, y = var_3624)[name = string("key_states_51")];
+            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([16])];
+            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([17])];
+            int32 concat_110_axis_0 = const()[name = string("concat_110_axis_0"), val = int32(0)];
+            bool concat_110_interleave_0 = const()[name = string("concat_110_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_110 = concat(axis = concat_110_axis_0, interleave = concat_110_interleave_0, values = (expand_dims_72, expand_dims_73, current_pos, expand_dims_75))[name = string("concat_110")];
+            tensor<int32, [1]> concat_111_values1_0 = const()[name = string("concat_111_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_111_values3_0 = const()[name = string("concat_111_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_111_axis_0 = const()[name = string("concat_111_axis_0"), val = int32(0)];
+            bool concat_111_interleave_0 = const()[name = string("concat_111_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_111 = concat(axis = concat_111_axis_0, interleave = concat_111_interleave_0, values = (expand_dims_76, concat_111_values1_0, var_616, concat_111_values3_0))[name = string("concat_111")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_13_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_13_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_13_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_110, begin_mask = model_model_kv_cache_0_internal_tensor_assign_13_begin_mask_0, end = concat_111, end_mask = model_model_kv_cache_0_internal_tensor_assign_13_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_13_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_13_stride_0, update = key_states_51, x = coreml_update_state_29)[name = string("model_model_kv_cache_0_internal_tensor_assign_13_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_13_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_30_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_30 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_30")];
+            tensor<int32, [1]> expand_dims_78 = const()[name = string("expand_dims_78"), val = tensor<int32, [1]>([44])];
+            tensor<int32, [1]> expand_dims_79 = const()[name = string("expand_dims_79"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_81 = const()[name = string("expand_dims_81"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_82 = const()[name = string("expand_dims_82"), val = tensor<int32, [1]>([45])];
+            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
+            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_78, expand_dims_79, current_pos, expand_dims_81))[name = string("concat_114")];
+            tensor<int32, [1]> concat_115_values1_0 = const()[name = string("concat_115_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_115_values3_0 = const()[name = string("concat_115_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_115_axis_0 = const()[name = string("concat_115_axis_0"), val = int32(0)];
+            bool concat_115_interleave_0 = const()[name = string("concat_115_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_115 = concat(axis = concat_115_axis_0, interleave = concat_115_interleave_0, values = (expand_dims_82, concat_115_values1_0, var_616, concat_115_values3_0))[name = string("concat_115")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_14_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_14_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_14_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_51 = transpose(perm = var_3557, x = var_3552)[name = string("transpose_23")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_114, begin_mask = model_model_kv_cache_0_internal_tensor_assign_14_begin_mask_0, end = concat_115, end_mask = model_model_kv_cache_0_internal_tensor_assign_14_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_14_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_14_stride_0, update = value_states_51, x = coreml_update_state_30)[name = string("model_model_kv_cache_0_internal_tensor_assign_14_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_14_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_31_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_31 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_31")];
+            tensor<int32, [4]> var_3695_begin_0 = const()[name = string("op_3695_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> var_3695_end_0 = const()[name = string("op_3695_end_0"), val = tensor<int32, [4]>([17, 2, 2048, 128])];
+            tensor<bool, [4]> var_3695_end_mask_0 = const()[name = string("op_3695_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3695_cast_fp16 = slice_by_index(begin = var_3695_begin_0, end = var_3695_end_0, end_mask = var_3695_end_mask_0, x = coreml_update_state_31)[name = string("op_3695_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_13_axes_0 = const()[name = string("K_layer_cache_13_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_13_cast_fp16 = squeeze(axes = K_layer_cache_13_axes_0, x = var_3695_cast_fp16)[name = string("K_layer_cache_13_cast_fp16")];
+            tensor<int32, [4]> var_3702_begin_0 = const()[name = string("op_3702_begin_0"), val = tensor<int32, [4]>([44, 0, 0, 0])];
+            tensor<int32, [4]> var_3702_end_0 = const()[name = string("op_3702_end_0"), val = tensor<int32, [4]>([45, 2, 2048, 128])];
+            tensor<bool, [4]> var_3702_end_mask_0 = const()[name = string("op_3702_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_3702_cast_fp16 = slice_by_index(begin = var_3702_begin_0, end = var_3702_end_0, end_mask = var_3702_end_mask_0, x = coreml_update_state_31)[name = string("op_3702_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_13_axes_0 = const()[name = string("V_layer_cache_13_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_13_cast_fp16 = squeeze(axes = V_layer_cache_13_axes_0, x = var_3702_cast_fp16)[name = string("V_layer_cache_13_cast_fp16")];
+            tensor<int32, [1]> x_99_axes_0 = const()[name = string("x_99_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_99_cast_fp16 = expand_dims(axes = x_99_axes_0, x = K_layer_cache_13_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [4]> var_3731 = const()[name = string("op_3731"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_101_cast_fp16 = tile(reps = var_3731, x = x_99_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [4]> var_3743 = const()[name = string("op_3743"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_55_cast_fp16 = reshape(shape = var_3743, x = x_101_cast_fp16)[name = string("key_states_55_cast_fp16")];
+            tensor<int32, [1]> x_105_axes_0 = const()[name = string("x_105_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_105_cast_fp16 = expand_dims(axes = x_105_axes_0, x = V_layer_cache_13_cast_fp16)[name = string("x_105_cast_fp16")];
+            tensor<int32, [4]> var_3751 = const()[name = string("op_3751"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_107_cast_fp16 = tile(reps = var_3751, x = x_105_cast_fp16)[name = string("x_107_cast_fp16")];
+            bool var_3786_transpose_x_1 = const()[name = string("op_3786_transpose_x_1"), val = bool(false)];
+            bool var_3786_transpose_y_1 = const()[name = string("op_3786_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_3786_cast_fp16 = matmul(transpose_x = var_3786_transpose_x_1, transpose_y = var_3786_transpose_y_1, x = query_states_39, y = key_states_55_cast_fp16)[name = string("op_3786_cast_fp16")];
+            fp16 var_3787_to_fp16 = const()[name = string("op_3787_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_25_cast_fp16 = mul(x = var_3786_cast_fp16, y = var_3787_to_fp16)[name = string("attn_logits_25_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_27_cast_fp16 = add(x = attn_logits_25_cast_fp16, y = causal_mask)[name = string("attn_logits_27_cast_fp16")];
+            int32 var_3814 = const()[name = string("op_3814"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_3816_cast_fp16 = softmax(axis = var_3814, x = attn_logits_27_cast_fp16)[name = string("op_3816_cast_fp16")];
+            tensor<int32, [3]> concat_120 = const()[name = string("concat_120"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_18_cast_fp16 = reshape(shape = concat_120, x = var_3816_cast_fp16)[name = string("reshape_18_cast_fp16")];
+            tensor<int32, [3]> concat_121 = const()[name = string("concat_121"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_19_cast_fp16 = reshape(shape = concat_121, x = x_107_cast_fp16)[name = string("reshape_19_cast_fp16")];
+            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
+            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = reshape_18_cast_fp16, y = reshape_19_cast_fp16)[name = string("matmul_6_cast_fp16")];
+            tensor<int32, [4]> concat_125 = const()[name = string("concat_125"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_20_cast_fp16 = reshape(shape = concat_125, x = matmul_6_cast_fp16)[name = string("reshape_20_cast_fp16")];
+            tensor<int32, [4]> var_3843_perm_0 = const()[name = string("op_3843_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3862 = const()[name = string("op_3862"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_3843 = transpose(perm = var_3843_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_22")];
+            tensor<fp16, [1, 64, 1536]> attn_output_65 = reshape(shape = var_3862, x = var_3843)[name = string("attn_output_65")];
+            tensor<int32, [3]> var_3867 = const()[name = string("op_3867"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_6_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319214784))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320984320))))[name = string("squeeze_6_palettized")];
+            string var_3883_pad_type_0 = const()[name = string("op_3883_pad_type_0"), val = string("valid")];
+            int32 var_3883_groups_0 = const()[name = string("op_3883_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_3883_strides_0 = const()[name = string("op_3883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_3883_pad_0 = const()[name = string("op_3883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_3883_dilations_0 = const()[name = string("op_3883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_3868 = transpose(perm = var_3867, x = attn_output_65)[name = string("transpose_21")];
+            tensor<fp16, [1, 1536, 64]> var_3883 = conv(dilations = var_3883_dilations_0, groups = var_3883_groups_0, pad = var_3883_pad_0, pad_type = var_3883_pad_type_0, strides = var_3883_strides_0, weight = squeeze_6_palettized, x = var_3868)[name = string("op_3883")];
+            tensor<int32, [3]> var_3887 = const()[name = string("op_3887"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_69 = transpose(perm = var_3887, x = var_3883)[name = string("transpose_20")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_41_cast_fp16 = add(x = hidden_states_37_cast_fp16, y = attn_output_69)[name = string("hidden_states_41_cast_fp16")];
+            int32 var_3900 = const()[name = string("op_3900"), val = int32(-1)];
+            fp16 const_179_promoted_to_fp16 = const()[name = string("const_179_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_3902_cast_fp16 = mul(x = hidden_states_41_cast_fp16, y = const_179_promoted_to_fp16)[name = string("op_3902_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_91_cast_fp16 = concat(axis = var_3900, interleave = input_91_interleave_0, values = (hidden_states_41_cast_fp16, var_3902_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [1]> normed_53_axes_0 = const()[name = string("normed_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3897_to_fp16 = const()[name = string("op_3897_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_53_cast_fp16 = layer_norm(axes = normed_53_axes_0, epsilon = var_3897_to_fp16, x = input_91_cast_fp16)[name = string("normed_53_cast_fp16")];
+            tensor<int32, [3]> normed_55_begin_0 = const()[name = string("normed_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_55_end_0 = const()[name = string("normed_55_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_55_end_mask_0 = const()[name = string("normed_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_55_cast_fp16 = slice_by_index(begin = normed_55_begin_0, end = normed_55_end_0, end_mask = normed_55_end_mask_0, x = normed_53_cast_fp16)[name = string("normed_55_cast_fp16")];
+            tensor<fp16, [1536]> const_182_promoted_to_fp16 = const()[name = string("const_182_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321033536)))];
+            tensor<fp16, [1, 64, 1536]> x_109_cast_fp16 = mul(x = normed_55_cast_fp16, y = const_182_promoted_to_fp16)[name = string("x_109_cast_fp16")];
+            tensor<int32, [3]> var_3927 = const()[name = string("op_3927"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_93_axes_0 = const()[name = string("input_93_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_3928 = transpose(perm = var_3927, x = x_109_cast_fp16)[name = string("transpose_19")];
+            tensor<fp16, [1, 1536, 1, 64]> input_93 = expand_dims(axes = input_93_axes_0, x = var_3928)[name = string("input_93")];
+            string input_95_pad_type_0 = const()[name = string("input_95_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_95_strides_0 = const()[name = string("input_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_95_pad_0 = const()[name = string("input_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_95_dilations_0 = const()[name = string("input_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_95_groups_0 = const()[name = string("input_95_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_95 = conv(dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = model_model_layers_16_mlp_gate_proj_weight_palettized, x = input_93)[name = string("input_95")];
+            string b_13_pad_type_0 = const()[name = string("b_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_13_strides_0 = const()[name = string("b_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_13_pad_0 = const()[name = string("b_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_13_dilations_0 = const()[name = string("b_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_13_groups_0 = const()[name = string("b_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_13 = conv(dilations = b_13_dilations_0, groups = b_13_groups_0, pad = b_13_pad_0, pad_type = b_13_pad_type_0, strides = b_13_strides_0, weight = model_model_layers_16_mlp_up_proj_weight_palettized, x = input_93)[name = string("b_13")];
+            tensor<fp16, [1, 8960, 1, 64]> c_13 = silu(x = input_95)[name = string("c_13")];
+            tensor<fp16, [1, 8960, 1, 64]> input_97 = mul(x = c_13, y = b_13)[name = string("input_97")];
+            string e_13_pad_type_0 = const()[name = string("e_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_13_strides_0 = const()[name = string("e_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_13_pad_0 = const()[name = string("e_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_13_dilations_0 = const()[name = string("e_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_13_groups_0 = const()[name = string("e_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_13 = conv(dilations = e_13_dilations_0, groups = e_13_groups_0, pad = e_13_pad_0, pad_type = e_13_pad_type_0, strides = e_13_strides_0, weight = model_model_layers_16_mlp_down_proj_weight_palettized, x = input_97)[name = string("e_13")];
+            tensor<int32, [1]> var_3950_axes_0 = const()[name = string("op_3950_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_3950 = squeeze(axes = var_3950_axes_0, x = e_13)[name = string("op_3950")];
+            tensor<int32, [3]> var_3951 = const()[name = string("op_3951"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_3952 = transpose(perm = var_3951, x = var_3950)[name = string("transpose_18")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_43_cast_fp16 = add(x = hidden_states_41_cast_fp16, y = var_3952)[name = string("hidden_states_43_cast_fp16")];
+            int32 var_3964 = const()[name = string("op_3964"), val = int32(-1)];
+            fp16 const_183_promoted_to_fp16 = const()[name = string("const_183_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_3966_cast_fp16 = mul(x = hidden_states_43_cast_fp16, y = const_183_promoted_to_fp16)[name = string("op_3966_cast_fp16")];
+            bool input_99_interleave_0 = const()[name = string("input_99_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_99_cast_fp16 = concat(axis = var_3964, interleave = input_99_interleave_0, values = (hidden_states_43_cast_fp16, var_3966_cast_fp16))[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> normed_57_axes_0 = const()[name = string("normed_57_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3961_to_fp16 = const()[name = string("op_3961_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_57_cast_fp16 = layer_norm(axes = normed_57_axes_0, epsilon = var_3961_to_fp16, x = input_99_cast_fp16)[name = string("normed_57_cast_fp16")];
+            tensor<int32, [3]> normed_59_begin_0 = const()[name = string("normed_59_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_59_end_0 = const()[name = string("normed_59_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_59_end_mask_0 = const()[name = string("normed_59_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_59_cast_fp16 = slice_by_index(begin = normed_59_begin_0, end = normed_59_end_0, end_mask = normed_59_end_mask_0, x = normed_57_cast_fp16)[name = string("normed_59_cast_fp16")];
+            tensor<fp16, [1536]> const_186_promoted_to_fp16 = const()[name = string("const_186_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321036672)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_45_cast_fp16 = mul(x = normed_59_cast_fp16, y = const_186_promoted_to_fp16)[name = string("hidden_states_45_cast_fp16")];
+            tensor<int32, [3]> var_3989 = const()[name = string("op_3989"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_3992_axes_0 = const()[name = string("op_3992_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_3990_cast_fp16 = transpose(perm = var_3989, x = hidden_states_45_cast_fp16)[name = string("transpose_17")];
+            tensor<fp16, [1, 1536, 1, 64]> var_3992_cast_fp16 = expand_dims(axes = var_3992_axes_0, x = var_3990_cast_fp16)[name = string("op_3992_cast_fp16")];
+            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_43 = conv(bias = model_model_layers_17_self_attn_q_proj_bias, dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = model_model_layers_17_self_attn_q_proj_weight_palettized, x = var_3992_cast_fp16)[name = string("query_states_43")];
+            string key_states_57_pad_type_0 = const()[name = string("key_states_57_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_57_strides_0 = const()[name = string("key_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_57_pad_0 = const()[name = string("key_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_57_dilations_0 = const()[name = string("key_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_57_groups_0 = const()[name = string("key_states_57_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_57 = conv(bias = model_model_layers_17_self_attn_k_proj_bias, dilations = key_states_57_dilations_0, groups = key_states_57_groups_0, pad = key_states_57_pad_0, pad_type = key_states_57_pad_type_0, strides = key_states_57_strides_0, weight = model_model_layers_17_self_attn_k_proj_weight_palettized, x = var_3992_cast_fp16)[name = string("key_states_57")];
+            string value_states_57_pad_type_0 = const()[name = string("value_states_57_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_57_strides_0 = const()[name = string("value_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_57_pad_0 = const()[name = string("value_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_57_dilations_0 = const()[name = string("value_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_57_groups_0 = const()[name = string("value_states_57_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_57 = conv(bias = model_model_layers_17_self_attn_v_proj_bias, dilations = value_states_57_dilations_0, groups = value_states_57_groups_0, pad = value_states_57_pad_0, pad_type = value_states_57_pad_type_0, strides = value_states_57_strides_0, weight = model_model_layers_17_self_attn_v_proj_weight_palettized, x = var_3992_cast_fp16)[name = string("value_states_57")];
+            tensor<int32, [4]> var_4034 = const()[name = string("op_4034"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_4035 = reshape(shape = var_4034, x = query_states_43)[name = string("op_4035")];
+            tensor<int32, [4]> var_4040 = const()[name = string("op_4040"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_4045 = const()[name = string("op_4045"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_4046 = reshape(shape = var_4045, x = key_states_57)[name = string("op_4046")];
+            tensor<int32, [4]> var_4051 = const()[name = string("op_4051"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_4056 = const()[name = string("op_4056"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_4057 = reshape(shape = var_4056, x = value_states_57)[name = string("op_4057")];
+            tensor<int32, [4]> var_4062 = const()[name = string("op_4062"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_29 = transpose(perm = var_4040, x = var_4035)[name = string("transpose_16")];
+            tensor<fp16, [1, 12, 64, 128]> var_4076 = mul(x = q_29, y = cos_5)[name = string("op_4076")];
+            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_29 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_29)[name = string("x1_29")];
+            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_29 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_29)[name = string("x2_29")];
+            fp16 const_190_promoted = const()[name = string("const_190_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_4097 = mul(x = x2_29, y = const_190_promoted)[name = string("op_4097")];
+            int32 var_4099 = const()[name = string("op_4099"), val = int32(-1)];
+            bool var_4100_interleave_0 = const()[name = string("op_4100_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_4100 = concat(axis = var_4099, interleave = var_4100_interleave_0, values = (var_4097, x1_29))[name = string("op_4100")];
+            tensor<fp16, [1, 12, 64, 128]> var_4101 = mul(x = var_4100, y = sin_5)[name = string("op_4101")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_45 = add(x = var_4076, y = var_4101)[name = string("query_states_45")];
+            tensor<fp16, [1, 2, 64, 128]> k_29 = transpose(perm = var_4051, x = var_4046)[name = string("transpose_15")];
+            tensor<fp16, [1, 2, 64, 128]> var_4104 = mul(x = k_29, y = cos_5)[name = string("op_4104")];
+            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1_31 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_29)[name = string("x1_31")];
+            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2_31 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_29)[name = string("x2_31")];
+            fp16 const_193_promoted = const()[name = string("const_193_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_4125 = mul(x = x2_31, y = const_193_promoted)[name = string("op_4125")];
+            int32 var_4127 = const()[name = string("op_4127"), val = int32(-1)];
+            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_4128 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125, x1_31))[name = string("op_4128")];
+            tensor<fp16, [1, 2, 64, 128]> var_4129 = mul(x = var_4128, y = sin_5)[name = string("op_4129")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_59 = add(x = var_4104, y = var_4129)[name = string("key_states_59")];
+            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([17])];
+            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([18])];
+            int32 concat_128_axis_0 = const()[name = string("concat_128_axis_0"), val = int32(0)];
+            bool concat_128_interleave_0 = const()[name = string("concat_128_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_128 = concat(axis = concat_128_axis_0, interleave = concat_128_interleave_0, values = (expand_dims_84, expand_dims_85, current_pos, expand_dims_87))[name = string("concat_128")];
+            tensor<int32, [1]> concat_129_values1_0 = const()[name = string("concat_129_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_129_values3_0 = const()[name = string("concat_129_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_129_axis_0 = const()[name = string("concat_129_axis_0"), val = int32(0)];
+            bool concat_129_interleave_0 = const()[name = string("concat_129_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_129 = concat(axis = concat_129_axis_0, interleave = concat_129_interleave_0, values = (expand_dims_88, concat_129_values1_0, var_616, concat_129_values3_0))[name = string("concat_129")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_15_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_15_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_15_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_128, begin_mask = model_model_kv_cache_0_internal_tensor_assign_15_begin_mask_0, end = concat_129, end_mask = model_model_kv_cache_0_internal_tensor_assign_15_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_15_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_15_stride_0, update = key_states_59, x = coreml_update_state_31)[name = string("model_model_kv_cache_0_internal_tensor_assign_15_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_15_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_32_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_32 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_32")];
+            tensor<int32, [1]> expand_dims_90 = const()[name = string("expand_dims_90"), val = tensor<int32, [1]>([45])];
+            tensor<int32, [1]> expand_dims_91 = const()[name = string("expand_dims_91"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_93 = const()[name = string("expand_dims_93"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_94 = const()[name = string("expand_dims_94"), val = tensor<int32, [1]>([46])];
+            int32 concat_132_axis_0 = const()[name = string("concat_132_axis_0"), val = int32(0)];
+            bool concat_132_interleave_0 = const()[name = string("concat_132_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_132 = concat(axis = concat_132_axis_0, interleave = concat_132_interleave_0, values = (expand_dims_90, expand_dims_91, current_pos, expand_dims_93))[name = string("concat_132")];
+            tensor<int32, [1]> concat_133_values1_0 = const()[name = string("concat_133_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_133_values3_0 = const()[name = string("concat_133_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_133_axis_0 = const()[name = string("concat_133_axis_0"), val = int32(0)];
+            bool concat_133_interleave_0 = const()[name = string("concat_133_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_133 = concat(axis = concat_133_axis_0, interleave = concat_133_interleave_0, values = (expand_dims_94, concat_133_values1_0, var_616, concat_133_values3_0))[name = string("concat_133")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_16_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_16_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_16_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_59 = transpose(perm = var_4062, x = var_4057)[name = string("transpose_14")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_132, begin_mask = model_model_kv_cache_0_internal_tensor_assign_16_begin_mask_0, end = concat_133, end_mask = model_model_kv_cache_0_internal_tensor_assign_16_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_16_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_16_stride_0, update = value_states_59, x = coreml_update_state_32)[name = string("model_model_kv_cache_0_internal_tensor_assign_16_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_16_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_33_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_33 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_33")];
+            tensor<int32, [4]> var_4200_begin_0 = const()[name = string("op_4200_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> var_4200_end_0 = const()[name = string("op_4200_end_0"), val = tensor<int32, [4]>([18, 2, 2048, 128])];
+            tensor<bool, [4]> var_4200_end_mask_0 = const()[name = string("op_4200_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_4200_cast_fp16 = slice_by_index(begin = var_4200_begin_0, end = var_4200_end_0, end_mask = var_4200_end_mask_0, x = coreml_update_state_33)[name = string("op_4200_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_15_axes_0 = const()[name = string("K_layer_cache_15_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_15_cast_fp16 = squeeze(axes = K_layer_cache_15_axes_0, x = var_4200_cast_fp16)[name = string("K_layer_cache_15_cast_fp16")];
+            tensor<int32, [4]> var_4207_begin_0 = const()[name = string("op_4207_begin_0"), val = tensor<int32, [4]>([45, 0, 0, 0])];
+            tensor<int32, [4]> var_4207_end_0 = const()[name = string("op_4207_end_0"), val = tensor<int32, [4]>([46, 2, 2048, 128])];
+            tensor<bool, [4]> var_4207_end_mask_0 = const()[name = string("op_4207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_4207_cast_fp16 = slice_by_index(begin = var_4207_begin_0, end = var_4207_end_0, end_mask = var_4207_end_mask_0, x = coreml_update_state_33)[name = string("op_4207_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_15_axes_0 = const()[name = string("V_layer_cache_15_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_15_cast_fp16 = squeeze(axes = V_layer_cache_15_axes_0, x = var_4207_cast_fp16)[name = string("V_layer_cache_15_cast_fp16")];
+            tensor<int32, [1]> x_115_axes_0 = const()[name = string("x_115_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_115_cast_fp16 = expand_dims(axes = x_115_axes_0, x = K_layer_cache_15_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_4236 = const()[name = string("op_4236"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_117_cast_fp16 = tile(reps = var_4236, x = x_115_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_4248 = const()[name = string("op_4248"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_63_cast_fp16 = reshape(shape = var_4248, x = x_117_cast_fp16)[name = string("key_states_63_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_121_cast_fp16 = expand_dims(axes = x_121_axes_0, x = V_layer_cache_15_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [4]> var_4256 = const()[name = string("op_4256"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_123_cast_fp16 = tile(reps = var_4256, x = x_121_cast_fp16)[name = string("x_123_cast_fp16")];
+            bool var_4291_transpose_x_1 = const()[name = string("op_4291_transpose_x_1"), val = bool(false)];
+            bool var_4291_transpose_y_1 = const()[name = string("op_4291_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_4291_cast_fp16 = matmul(transpose_x = var_4291_transpose_x_1, transpose_y = var_4291_transpose_y_1, x = query_states_45, y = key_states_63_cast_fp16)[name = string("op_4291_cast_fp16")];
+            fp16 var_4292_to_fp16 = const()[name = string("op_4292_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_29_cast_fp16 = mul(x = var_4291_cast_fp16, y = var_4292_to_fp16)[name = string("attn_logits_29_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_31_cast_fp16 = add(x = attn_logits_29_cast_fp16, y = causal_mask)[name = string("attn_logits_31_cast_fp16")];
+            int32 var_4319 = const()[name = string("op_4319"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_4321_cast_fp16 = softmax(axis = var_4319, x = attn_logits_31_cast_fp16)[name = string("op_4321_cast_fp16")];
+            tensor<int32, [3]> concat_138 = const()[name = string("concat_138"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_21_cast_fp16 = reshape(shape = concat_138, x = var_4321_cast_fp16)[name = string("reshape_21_cast_fp16")];
+            tensor<int32, [3]> concat_139 = const()[name = string("concat_139"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_22_cast_fp16 = reshape(shape = concat_139, x = x_123_cast_fp16)[name = string("reshape_22_cast_fp16")];
+            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
+            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = reshape_21_cast_fp16, y = reshape_22_cast_fp16)[name = string("matmul_7_cast_fp16")];
+            tensor<int32, [4]> concat_143 = const()[name = string("concat_143"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_23_cast_fp16 = reshape(shape = concat_143, x = matmul_7_cast_fp16)[name = string("reshape_23_cast_fp16")];
+            tensor<int32, [4]> var_4348_perm_0 = const()[name = string("op_4348_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4367 = const()[name = string("op_4367"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_4348 = transpose(perm = var_4348_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_13")];
+            tensor<fp16, [1, 64, 1536]> attn_output_75 = reshape(shape = var_4367, x = var_4348)[name = string("attn_output_75")];
+            tensor<int32, [3]> var_4372 = const()[name = string("op_4372"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_7_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321039808))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322809344))))[name = string("squeeze_7_palettized")];
+            string var_4388_pad_type_0 = const()[name = string("op_4388_pad_type_0"), val = string("valid")];
+            int32 var_4388_groups_0 = const()[name = string("op_4388_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_4388_strides_0 = const()[name = string("op_4388_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_4388_pad_0 = const()[name = string("op_4388_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_4388_dilations_0 = const()[name = string("op_4388_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_4373 = transpose(perm = var_4372, x = attn_output_75)[name = string("transpose_12")];
+            tensor<fp16, [1, 1536, 64]> var_4388 = conv(dilations = var_4388_dilations_0, groups = var_4388_groups_0, pad = var_4388_pad_0, pad_type = var_4388_pad_type_0, strides = var_4388_strides_0, weight = squeeze_7_palettized, x = var_4373)[name = string("op_4388")];
+            tensor<int32, [3]> var_4392 = const()[name = string("op_4392"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output_79 = transpose(perm = var_4392, x = var_4388)[name = string("transpose_11")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_47_cast_fp16 = add(x = hidden_states_43_cast_fp16, y = attn_output_79)[name = string("hidden_states_47_cast_fp16")];
+            int32 var_4405 = const()[name = string("op_4405"), val = int32(-1)];
+            fp16 const_205_promoted_to_fp16 = const()[name = string("const_205_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_4407_cast_fp16 = mul(x = hidden_states_47_cast_fp16, y = const_205_promoted_to_fp16)[name = string("op_4407_cast_fp16")];
+            bool input_105_interleave_0 = const()[name = string("input_105_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_105_cast_fp16 = concat(axis = var_4405, interleave = input_105_interleave_0, values = (hidden_states_47_cast_fp16, var_4407_cast_fp16))[name = string("input_105_cast_fp16")];
+            tensor<int32, [1]> normed_61_axes_0 = const()[name = string("normed_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4402_to_fp16 = const()[name = string("op_4402_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_61_cast_fp16 = layer_norm(axes = normed_61_axes_0, epsilon = var_4402_to_fp16, x = input_105_cast_fp16)[name = string("normed_61_cast_fp16")];
+            tensor<int32, [3]> normed_63_begin_0 = const()[name = string("normed_63_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_63_end_0 = const()[name = string("normed_63_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_63_end_mask_0 = const()[name = string("normed_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_63_cast_fp16 = slice_by_index(begin = normed_63_begin_0, end = normed_63_end_0, end_mask = normed_63_end_mask_0, x = normed_61_cast_fp16)[name = string("normed_63_cast_fp16")];
+            tensor<fp16, [1536]> const_208_promoted_to_fp16 = const()[name = string("const_208_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322858560)))];
+            tensor<fp16, [1, 64, 1536]> x_125_cast_fp16 = mul(x = normed_63_cast_fp16, y = const_208_promoted_to_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> var_4432 = const()[name = string("op_4432"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_4433 = transpose(perm = var_4432, x = x_125_cast_fp16)[name = string("transpose_10")];
+            tensor<fp16, [1, 1536, 1, 64]> input_107 = expand_dims(axes = input_107_axes_0, x = var_4433)[name = string("input_107")];
+            string input_109_pad_type_0 = const()[name = string("input_109_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_109_strides_0 = const()[name = string("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_109_pad_0 = const()[name = string("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_109_dilations_0 = const()[name = string("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_109_groups_0 = const()[name = string("input_109_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_109 = conv(dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = model_model_layers_17_mlp_gate_proj_weight_palettized, x = input_107)[name = string("input_109")];
+            string b_15_pad_type_0 = const()[name = string("b_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_15_strides_0 = const()[name = string("b_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_15_pad_0 = const()[name = string("b_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_15_dilations_0 = const()[name = string("b_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_15_groups_0 = const()[name = string("b_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b_15 = conv(dilations = b_15_dilations_0, groups = b_15_groups_0, pad = b_15_pad_0, pad_type = b_15_pad_type_0, strides = b_15_strides_0, weight = model_model_layers_17_mlp_up_proj_weight_palettized, x = input_107)[name = string("b_15")];
+            tensor<fp16, [1, 8960, 1, 64]> c_15 = silu(x = input_109)[name = string("c_15")];
+            tensor<fp16, [1, 8960, 1, 64]> input_111 = mul(x = c_15, y = b_15)[name = string("input_111")];
+            string e_15_pad_type_0 = const()[name = string("e_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_15_strides_0 = const()[name = string("e_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_15_pad_0 = const()[name = string("e_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_15_dilations_0 = const()[name = string("e_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_15_groups_0 = const()[name = string("e_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e_15 = conv(dilations = e_15_dilations_0, groups = e_15_groups_0, pad = e_15_pad_0, pad_type = e_15_pad_type_0, strides = e_15_strides_0, weight = model_model_layers_17_mlp_down_proj_weight_palettized, x = input_111)[name = string("e_15")];
+            tensor<int32, [1]> var_4455_axes_0 = const()[name = string("op_4455_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_4455 = squeeze(axes = var_4455_axes_0, x = e_15)[name = string("op_4455")];
+            tensor<int32, [3]> var_4456 = const()[name = string("op_4456"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_4457 = transpose(perm = var_4456, x = var_4455)[name = string("transpose_9")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_49_cast_fp16 = add(x = hidden_states_47_cast_fp16, y = var_4457)[name = string("hidden_states_49_cast_fp16")];
+            int32 var_4469 = const()[name = string("op_4469"), val = int32(-1)];
+            fp16 const_209_promoted_to_fp16 = const()[name = string("const_209_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_4471_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_209_promoted_to_fp16)[name = string("op_4471_cast_fp16")];
+            bool input_113_interleave_0 = const()[name = string("input_113_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_113_cast_fp16 = concat(axis = var_4469, interleave = input_113_interleave_0, values = (hidden_states_49_cast_fp16, var_4471_cast_fp16))[name = string("input_113_cast_fp16")];
+            tensor<int32, [1]> normed_65_axes_0 = const()[name = string("normed_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4466_to_fp16 = const()[name = string("op_4466_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_65_cast_fp16 = layer_norm(axes = normed_65_axes_0, epsilon = var_4466_to_fp16, x = input_113_cast_fp16)[name = string("normed_65_cast_fp16")];
+            tensor<int32, [3]> normed_67_begin_0 = const()[name = string("normed_67_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_67_end_0 = const()[name = string("normed_67_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_67_end_mask_0 = const()[name = string("normed_67_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_67_cast_fp16 = slice_by_index(begin = normed_67_begin_0, end = normed_67_end_0, end_mask = normed_67_end_mask_0, x = normed_65_cast_fp16)[name = string("normed_67_cast_fp16")];
+            tensor<fp16, [1536]> const_212_promoted_to_fp16 = const()[name = string("const_212_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322861696)))];
+            tensor<fp16, [1, 64, 1536]> hidden_states_51_cast_fp16 = mul(x = normed_67_cast_fp16, y = const_212_promoted_to_fp16)[name = string("hidden_states_51_cast_fp16")];
+            tensor<int32, [3]> var_4494 = const()[name = string("op_4494"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_4497_axes_0 = const()[name = string("op_4497_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_4495_cast_fp16 = transpose(perm = var_4494, x = hidden_states_51_cast_fp16)[name = string("transpose_8")];
+            tensor<fp16, [1, 1536, 1, 64]> var_4497_cast_fp16 = expand_dims(axes = var_4497_axes_0, x = var_4495_cast_fp16)[name = string("op_4497_cast_fp16")];
+            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> query_states_49 = conv(bias = model_model_layers_18_self_attn_q_proj_bias, dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = model_model_layers_18_self_attn_q_proj_weight_palettized, x = var_4497_cast_fp16)[name = string("query_states_49")];
+            string key_states_65_pad_type_0 = const()[name = string("key_states_65_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> key_states_65_strides_0 = const()[name = string("key_states_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> key_states_65_pad_0 = const()[name = string("key_states_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> key_states_65_dilations_0 = const()[name = string("key_states_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 key_states_65_groups_0 = const()[name = string("key_states_65_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> key_states_65 = conv(bias = model_model_layers_18_self_attn_k_proj_bias, dilations = key_states_65_dilations_0, groups = key_states_65_groups_0, pad = key_states_65_pad_0, pad_type = key_states_65_pad_type_0, strides = key_states_65_strides_0, weight = model_model_layers_18_self_attn_k_proj_weight_palettized, x = var_4497_cast_fp16)[name = string("key_states_65")];
+            string value_states_65_pad_type_0 = const()[name = string("value_states_65_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> value_states_65_strides_0 = const()[name = string("value_states_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> value_states_65_pad_0 = const()[name = string("value_states_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> value_states_65_dilations_0 = const()[name = string("value_states_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 value_states_65_groups_0 = const()[name = string("value_states_65_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 64]> value_states_65 = conv(bias = model_model_layers_18_self_attn_v_proj_bias, dilations = value_states_65_dilations_0, groups = value_states_65_groups_0, pad = value_states_65_pad_0, pad_type = value_states_65_pad_type_0, strides = value_states_65_strides_0, weight = model_model_layers_18_self_attn_v_proj_weight_palettized, x = var_4497_cast_fp16)[name = string("value_states_65")];
+            tensor<int32, [4]> var_4539 = const()[name = string("op_4539"), val = tensor<int32, [4]>([1, 12, 128, 64])];
+            tensor<fp16, [1, 12, 128, 64]> var_4540 = reshape(shape = var_4539, x = query_states_49)[name = string("op_4540")];
+            tensor<int32, [4]> var_4545 = const()[name = string("op_4545"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_4550 = const()[name = string("op_4550"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_4551 = reshape(shape = var_4550, x = key_states_65)[name = string("op_4551")];
+            tensor<int32, [4]> var_4556 = const()[name = string("op_4556"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [4]> var_4561 = const()[name = string("op_4561"), val = tensor<int32, [4]>([1, 2, 128, 64])];
+            tensor<fp16, [1, 2, 128, 64]> var_4562 = reshape(shape = var_4561, x = value_states_65)[name = string("op_4562")];
+            tensor<int32, [4]> var_4567 = const()[name = string("op_4567"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<fp16, [1, 12, 64, 128]> q_33 = transpose(perm = var_4545, x = var_4540)[name = string("transpose_7")];
+            tensor<fp16, [1, 12, 64, 128]> var_4581 = mul(x = q_33, y = cos_5)[name = string("op_4581")];
+            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 12, 64, 64])];
+            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 12, 64, 64]> x1_33 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_33)[name = string("x1_33")];
+            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 12, 64, 64]> x2_33 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_33)[name = string("x2_33")];
+            fp16 const_216_promoted = const()[name = string("const_216_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 12, 64, 64]> var_4602 = mul(x = x2_33, y = const_216_promoted)[name = string("op_4602")];
+            int32 var_4604 = const()[name = string("op_4604"), val = int32(-1)];
+            bool var_4605_interleave_0 = const()[name = string("op_4605_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 64, 128]> var_4605 = concat(axis = var_4604, interleave = var_4605_interleave_0, values = (var_4602, x1_33))[name = string("op_4605")];
+            tensor<fp16, [1, 12, 64, 128]> var_4606 = mul(x = var_4605, y = sin_5)[name = string("op_4606")];
+            tensor<fp16, [1, 12, 64, 128]> query_states_51 = add(x = var_4581, y = var_4606)[name = string("query_states_51")];
+            tensor<fp16, [1, 2, 64, 128]> k_33 = transpose(perm = var_4556, x = var_4551)[name = string("transpose_6")];
+            tensor<fp16, [1, 2, 64, 128]> var_4609 = mul(x = k_33, y = cos_5)[name = string("op_4609")];
+            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
+            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 2, 64, 64]> x1 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_33)[name = string("x1")];
+            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
+            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 2, 64, 64]> x2 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_33)[name = string("x2")];
+            fp16 const_219_promoted = const()[name = string("const_219_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 2, 64, 64]> var_4630 = mul(x = x2, y = const_219_promoted)[name = string("op_4630")];
+            int32 var_4632 = const()[name = string("op_4632"), val = int32(-1)];
+            bool var_4633_interleave_0 = const()[name = string("op_4633_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2, 64, 128]> var_4633 = concat(axis = var_4632, interleave = var_4633_interleave_0, values = (var_4630, x1))[name = string("op_4633")];
+            tensor<fp16, [1, 2, 64, 128]> var_4634 = mul(x = var_4633, y = sin_5)[name = string("op_4634")];
+            tensor<fp16, [1, 2, 64, 128]> key_states_67 = add(x = var_4609, y = var_4634)[name = string("key_states_67")];
+            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([18])];
+            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([19])];
+            int32 concat_146_axis_0 = const()[name = string("concat_146_axis_0"), val = int32(0)];
+            bool concat_146_interleave_0 = const()[name = string("concat_146_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_146 = concat(axis = concat_146_axis_0, interleave = concat_146_interleave_0, values = (expand_dims_96, expand_dims_97, current_pos, expand_dims_99))[name = string("concat_146")];
+            tensor<int32, [1]> concat_147_values1_0 = const()[name = string("concat_147_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_147_values3_0 = const()[name = string("concat_147_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_147_axis_0 = const()[name = string("concat_147_axis_0"), val = int32(0)];
+            bool concat_147_interleave_0 = const()[name = string("concat_147_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_147 = concat(axis = concat_147_axis_0, interleave = concat_147_interleave_0, values = (expand_dims_100, concat_147_values1_0, var_616, concat_147_values3_0))[name = string("concat_147")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_17_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_17_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_17_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_146, begin_mask = model_model_kv_cache_0_internal_tensor_assign_17_begin_mask_0, end = concat_147, end_mask = model_model_kv_cache_0_internal_tensor_assign_17_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_17_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_17_stride_0, update = key_states_67, x = coreml_update_state_33)[name = string("model_model_kv_cache_0_internal_tensor_assign_17_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_17_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_34_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_34 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_34")];
+            tensor<int32, [1]> expand_dims_102 = const()[name = string("expand_dims_102"), val = tensor<int32, [1]>([46])];
+            tensor<int32, [1]> expand_dims_103 = const()[name = string("expand_dims_103"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_105 = const()[name = string("expand_dims_105"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> expand_dims_106 = const()[name = string("expand_dims_106"), val = tensor<int32, [1]>([47])];
+            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
+            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_102, expand_dims_103, current_pos, expand_dims_105))[name = string("concat_150")];
+            tensor<int32, [1]> concat_151_values1_0 = const()[name = string("concat_151_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_151_values3_0 = const()[name = string("concat_151_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_151_axis_0 = const()[name = string("concat_151_axis_0"), val = int32(0)];
+            bool concat_151_interleave_0 = const()[name = string("concat_151_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_151 = concat(axis = concat_151_axis_0, interleave = concat_151_interleave_0, values = (expand_dims_106, concat_151_values1_0, var_616, concat_151_values3_0))[name = string("concat_151")];
+            tensor<int32, [4]> model_model_kv_cache_0_internal_tensor_assign_18_stride_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_18_begin_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_18_end_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> model_model_kv_cache_0_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("model_model_kv_cache_0_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [1, 2, 64, 128]> value_states_67 = transpose(perm = var_4567, x = var_4562)[name = string("transpose_5")];
+            tensor<fp16, [56, 2, 2048, 128]> model_model_kv_cache_0_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_150, begin_mask = model_model_kv_cache_0_internal_tensor_assign_18_begin_mask_0, end = concat_151, end_mask = model_model_kv_cache_0_internal_tensor_assign_18_end_mask_0, squeeze_mask = model_model_kv_cache_0_internal_tensor_assign_18_squeeze_mask_0, stride = model_model_kv_cache_0_internal_tensor_assign_18_stride_0, update = value_states_67, x = coreml_update_state_34)[name = string("model_model_kv_cache_0_internal_tensor_assign_18_cast_fp16")];
+            write_state(data = model_model_kv_cache_0_internal_tensor_assign_18_cast_fp16, input = model_model_kv_cache_0)[name = string("coreml_update_state_35_write_state")];
+            tensor<fp16, [56, 2, 2048, 128]> coreml_update_state_35 = read_state(input = model_model_kv_cache_0)[name = string("coreml_update_state_35")];
+            tensor<int32, [4]> var_4705_begin_0 = const()[name = string("op_4705_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> var_4705_end_0 = const()[name = string("op_4705_end_0"), val = tensor<int32, [4]>([19, 2, 2048, 128])];
+            tensor<bool, [4]> var_4705_end_mask_0 = const()[name = string("op_4705_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_4705_cast_fp16 = slice_by_index(begin = var_4705_begin_0, end = var_4705_end_0, end_mask = var_4705_end_mask_0, x = coreml_update_state_35)[name = string("op_4705_cast_fp16")];
+            tensor<int32, [1]> K_layer_cache_axes_0 = const()[name = string("K_layer_cache_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> K_layer_cache_cast_fp16 = squeeze(axes = K_layer_cache_axes_0, x = var_4705_cast_fp16)[name = string("K_layer_cache_cast_fp16")];
+            tensor<int32, [4]> var_4712_begin_0 = const()[name = string("op_4712_begin_0"), val = tensor<int32, [4]>([46, 0, 0, 0])];
+            tensor<int32, [4]> var_4712_end_0 = const()[name = string("op_4712_end_0"), val = tensor<int32, [4]>([47, 2, 2048, 128])];
+            tensor<bool, [4]> var_4712_end_mask_0 = const()[name = string("op_4712_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 2, 2048, 128]> var_4712_cast_fp16 = slice_by_index(begin = var_4712_begin_0, end = var_4712_end_0, end_mask = var_4712_end_mask_0, x = coreml_update_state_35)[name = string("op_4712_cast_fp16")];
+            tensor<int32, [1]> V_layer_cache_axes_0 = const()[name = string("V_layer_cache_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [2, 2048, 128]> V_layer_cache_cast_fp16 = squeeze(axes = V_layer_cache_axes_0, x = var_4712_cast_fp16)[name = string("V_layer_cache_cast_fp16")];
+            tensor<int32, [1]> x_131_axes_0 = const()[name = string("x_131_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_131_cast_fp16 = expand_dims(axes = x_131_axes_0, x = K_layer_cache_cast_fp16)[name = string("x_131_cast_fp16")];
+            tensor<int32, [4]> var_4741 = const()[name = string("op_4741"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_133_cast_fp16 = tile(reps = var_4741, x = x_131_cast_fp16)[name = string("x_133_cast_fp16")];
+            tensor<int32, [4]> var_4753 = const()[name = string("op_4753"), val = tensor<int32, [4]>([1, -1, 2048, 128])];
+            tensor<fp16, [1, 12, 2048, 128]> key_states_cast_fp16 = reshape(shape = var_4753, x = x_133_cast_fp16)[name = string("key_states_cast_fp16")];
+            tensor<int32, [1]> x_137_axes_0 = const()[name = string("x_137_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 2048, 128]> x_137_cast_fp16 = expand_dims(axes = x_137_axes_0, x = V_layer_cache_cast_fp16)[name = string("x_137_cast_fp16")];
+            tensor<int32, [4]> var_4761 = const()[name = string("op_4761"), val = tensor<int32, [4]>([1, 6, 1, 1])];
+            tensor<fp16, [2, 6, 2048, 128]> x_139_cast_fp16 = tile(reps = var_4761, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            bool var_4796_transpose_x_1 = const()[name = string("op_4796_transpose_x_1"), val = bool(false)];
+            bool var_4796_transpose_y_1 = const()[name = string("op_4796_transpose_y_1"), val = bool(true)];
+            tensor<fp16, [1, 12, 64, 2048]> var_4796_cast_fp16 = matmul(transpose_x = var_4796_transpose_x_1, transpose_y = var_4796_transpose_y_1, x = query_states_51, y = key_states_cast_fp16)[name = string("op_4796_cast_fp16")];
+            fp16 var_4797_to_fp16 = const()[name = string("op_4797_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_33_cast_fp16 = mul(x = var_4796_cast_fp16, y = var_4797_to_fp16)[name = string("attn_logits_33_cast_fp16")];
+            tensor<fp16, [1, 12, 64, 2048]> attn_logits_cast_fp16 = add(x = attn_logits_33_cast_fp16, y = causal_mask)[name = string("attn_logits_cast_fp16")];
+            int32 var_4824 = const()[name = string("op_4824"), val = int32(-1)];
+            tensor<fp16, [1, 12, 64, 2048]> var_4826_cast_fp16 = softmax(axis = var_4824, x = attn_logits_cast_fp16)[name = string("op_4826_cast_fp16")];
+            tensor<int32, [3]> concat_156 = const()[name = string("concat_156"), val = tensor<int32, [3]>([12, 64, 2048])];
+            tensor<fp16, [12, 64, 2048]> reshape_24_cast_fp16 = reshape(shape = concat_156, x = var_4826_cast_fp16)[name = string("reshape_24_cast_fp16")];
+            tensor<int32, [3]> concat_157 = const()[name = string("concat_157"), val = tensor<int32, [3]>([12, 2048, 128])];
+            tensor<fp16, [12, 2048, 128]> reshape_25_cast_fp16 = reshape(shape = concat_157, x = x_139_cast_fp16)[name = string("reshape_25_cast_fp16")];
+            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
+            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [12, 64, 128]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = reshape_24_cast_fp16, y = reshape_25_cast_fp16)[name = string("matmul_8_cast_fp16")];
+            tensor<int32, [4]> concat_161 = const()[name = string("concat_161"), val = tensor<int32, [4]>([1, 12, 64, 128])];
+            tensor<fp16, [1, 12, 64, 128]> reshape_26_cast_fp16 = reshape(shape = concat_161, x = matmul_8_cast_fp16)[name = string("reshape_26_cast_fp16")];
+            tensor<int32, [4]> var_4853_perm_0 = const()[name = string("op_4853_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4872 = const()[name = string("op_4872"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<fp16, [1, 64, 12, 128]> var_4853 = transpose(perm = var_4853_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_4")];
+            tensor<fp16, [1, 64, 1536]> attn_output_85 = reshape(shape = var_4872, x = var_4853)[name = string("attn_output_85")];
+            tensor<int32, [3]> var_4877 = const()[name = string("op_4877"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 1536, 1]> squeeze_8_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322864832))), lut = tensor<fp16, [384, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324634368))))[name = string("squeeze_8_palettized")];
+            string var_4893_pad_type_0 = const()[name = string("op_4893_pad_type_0"), val = string("valid")];
+            int32 var_4893_groups_0 = const()[name = string("op_4893_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_4893_strides_0 = const()[name = string("op_4893_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_4893_pad_0 = const()[name = string("op_4893_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_4893_dilations_0 = const()[name = string("op_4893_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536, 64]> var_4878 = transpose(perm = var_4877, x = attn_output_85)[name = string("transpose_3")];
+            tensor<fp16, [1, 1536, 64]> var_4893 = conv(dilations = var_4893_dilations_0, groups = var_4893_groups_0, pad = var_4893_pad_0, pad_type = var_4893_pad_type_0, strides = var_4893_strides_0, weight = squeeze_8_palettized, x = var_4878)[name = string("op_4893")];
+            tensor<int32, [3]> var_4897 = const()[name = string("op_4897"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> attn_output = transpose(perm = var_4897, x = var_4893)[name = string("transpose_2")];
+            tensor<fp16, [1, 64, 1536]> hidden_states_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = attn_output)[name = string("hidden_states_cast_fp16")];
+            int32 var_4910 = const()[name = string("op_4910"), val = int32(-1)];
+            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 64, 1536]> var_4912_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_4912_cast_fp16")];
+            bool input_119_interleave_0 = const()[name = string("input_119_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 64, 3072]> input_119_cast_fp16 = concat(axis = var_4910, interleave = input_119_interleave_0, values = (hidden_states_cast_fp16, var_4912_cast_fp16))[name = string("input_119_cast_fp16")];
+            tensor<int32, [1]> normed_69_axes_0 = const()[name = string("normed_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4907_to_fp16 = const()[name = string("op_4907_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 64, 3072]> normed_69_cast_fp16 = layer_norm(axes = normed_69_axes_0, epsilon = var_4907_to_fp16, x = input_119_cast_fp16)[name = string("normed_69_cast_fp16")];
+            tensor<int32, [3]> normed_begin_0 = const()[name = string("normed_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> normed_end_0 = const()[name = string("normed_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
+            tensor<bool, [3]> normed_end_mask_0 = const()[name = string("normed_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 64, 1536]> normed_cast_fp16 = slice_by_index(begin = normed_begin_0, end = normed_end_0, end_mask = normed_end_mask_0, x = normed_69_cast_fp16)[name = string("normed_cast_fp16")];
+            tensor<fp16, [1536]> const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324683584)))];
+            tensor<fp16, [1, 64, 1536]> x_141_cast_fp16 = mul(x = normed_cast_fp16, y = const_234_promoted_to_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [3]> var_4937 = const()[name = string("op_4937"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_4938 = transpose(perm = var_4937, x = x_141_cast_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1536, 1, 64]> input_121 = expand_dims(axes = input_121_axes_0, x = var_4938)[name = string("input_121")];
+            string input_123_pad_type_0 = const()[name = string("input_123_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_123_strides_0 = const()[name = string("input_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_123_pad_0 = const()[name = string("input_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_123_dilations_0 = const()[name = string("input_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_123_groups_0 = const()[name = string("input_123_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> input_123 = conv(dilations = input_123_dilations_0, groups = input_123_groups_0, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = input_123_strides_0, weight = model_model_layers_18_mlp_gate_proj_weight_palettized, x = input_121)[name = string("input_123")];
+            string b_pad_type_0 = const()[name = string("b_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> b_strides_0 = const()[name = string("b_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> b_pad_0 = const()[name = string("b_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> b_dilations_0 = const()[name = string("b_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 b_groups_0 = const()[name = string("b_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 8960, 1, 64]> b = conv(dilations = b_dilations_0, groups = b_groups_0, pad = b_pad_0, pad_type = b_pad_type_0, strides = b_strides_0, weight = model_model_layers_18_mlp_up_proj_weight_palettized, x = input_121)[name = string("b")];
+            tensor<fp16, [1, 8960, 1, 64]> c = silu(x = input_123)[name = string("c")];
+            tensor<fp16, [1, 8960, 1, 64]> input = mul(x = c, y = b)[name = string("input")];
+            string e_pad_type_0 = const()[name = string("e_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> e_strides_0 = const()[name = string("e_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> e_pad_0 = const()[name = string("e_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> e_dilations_0 = const()[name = string("e_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 e_groups_0 = const()[name = string("e_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 64]> e = conv(dilations = e_dilations_0, groups = e_groups_0, pad = e_pad_0, pad_type = e_pad_type_0, strides = e_strides_0, weight = model_model_layers_18_mlp_down_proj_weight_palettized, x = input)[name = string("e")];
+            tensor<int32, [1]> var_4960_axes_0 = const()[name = string("op_4960_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 64]> var_4960 = squeeze(axes = var_4960_axes_0, x = e)[name = string("op_4960")];
+            tensor<int32, [3]> var_4961 = const()[name = string("op_4961"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 64, 1536]> var_4962 = transpose(perm = var_4961, x = var_4960)[name = string("transpose_0")];
+            tensor<fp16, [1, 64, 1536]> output_hidden_states = add(x = hidden_states_cast_fp16, y = var_4962)[name = string("op_4964_cast_fp16")];
+        } -> (output_hidden_states);
+}
\ No newline at end of file