diff --git "a/chunk1.mlmodelc/model.mil" "b/chunk1.mlmodelc/model.mil"
new file mode 100644--- /dev/null
+++ "b/chunk1.mlmodelc/model.mil"
@@ -0,0 +1,3940 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func main<ios18>(tensor<fp16, [1, 1, 1, 512]> causal_mask, tensor<fp16, [1, 1, 1536]> hidden_states, state<tensor<fp16, [24, 1, 512, 512]>> kv_cache_0, tensor<fp16, [1, 1, 8960]> per_layer_combined, tensor<int32, [1]> position_ids, tensor<fp16, [1, 1, 512, 1]> update_mask) {
+            tensor<fp16, [1024, 512]> sin_full_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), lut = tensor<fp16, [32, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262272))))[name = string("sin_full_palettized")];
+            tensor<fp16, [1024, 512]> cos_full_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263360))), lut = tensor<fp16, [32, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525568))))[name = string("cos_full_palettized")];
+            tensor<fp16, [1024, 256]> sin_sliding_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526656))), lut = tensor<fp16, [32, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657792))))[name = string("sin_sliding_palettized")];
+            tensor<fp16, [1024, 256]> cos_sliding_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658880))), lut = tensor<fp16, [32, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790016))))[name = string("cos_sliding_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_0_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791104))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2364032))))[name = string("layers_0_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_0_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2366144))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2562816))))[name = string("layers_0_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_0_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2563136))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2759808))))[name = string("layers_0_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_0_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2760128))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7478784))))[name = string("layers_0_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_0_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7484992))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12203648))))[name = string("layers_0_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_0_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12209856))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16928512))))[name = string("layers_0_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_0_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16930112))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17126784))))[name = string("layers_0_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_1_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17127104))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18700032))))[name = string("layers_1_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_1_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18702144))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18898816))))[name = string("layers_1_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_1_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18899136))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19095808))))[name = string("layers_1_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_1_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19096128))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23814784))))[name = string("layers_1_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_1_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23820992))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28539648))))[name = string("layers_1_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_1_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28545856))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33264512))))[name = string("layers_1_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_1_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33266112))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33462784))))[name = string("layers_1_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_2_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33463104))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35036032))))[name = string("layers_2_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_2_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35038144))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35234816))))[name = string("layers_2_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_2_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35235136))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35431808))))[name = string("layers_2_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_2_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35432128))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40150784))))[name = string("layers_2_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_2_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40156992))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44875648))))[name = string("layers_2_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_2_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44881856))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49600512))))[name = string("layers_2_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_2_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49602112))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49798784))))[name = string("layers_2_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_3_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49799104))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51372032))))[name = string("layers_3_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_3_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51374144))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51570816))))[name = string("layers_3_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_3_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51571136))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51767808))))[name = string("layers_3_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_3_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51768128))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56486784))))[name = string("layers_3_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_3_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56492992))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61211648))))[name = string("layers_3_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_3_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61217856))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65936512))))[name = string("layers_3_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_3_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65938112))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66134784))))[name = string("layers_3_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [4096, 1536, 1, 1]> layers_4_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4096, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66135104))), lut = tensor<fp16, [128, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69280896))))[name = string("layers_4_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [512, 1536, 1, 1]> layers_4_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [512, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69285056))), lut = tensor<fp16, [16, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69678336))))[name = string("layers_4_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [512, 1536, 1, 1]> layers_4_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [512, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69678912))), lut = tensor<fp16, [16, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70072192))))[name = string("layers_4_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_4_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70072768))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74791424))))[name = string("layers_4_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_4_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74797632))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79516288))))[name = string("layers_4_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_4_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79522496))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84241152))))[name = string("layers_4_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_4_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84242752))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84439424))))[name = string("layers_4_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_5_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84439744))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86012672))))[name = string("layers_5_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_5_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86014784))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86211456))))[name = string("layers_5_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_5_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86211776))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86408448))))[name = string("layers_5_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_5_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86408768))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91127424))))[name = string("layers_5_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_5_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91133632))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95852288))))[name = string("layers_5_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_5_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95858496))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100577152))))[name = string("layers_5_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_5_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100578752))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100775424))))[name = string("layers_5_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_6_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100775744))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102348672))))[name = string("layers_6_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_6_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102350784))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102547456))))[name = string("layers_6_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_6_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102547776))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102744448))))[name = string("layers_6_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_6_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102744768))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107463424))))[name = string("layers_6_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_6_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107469632))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112188288))))[name = string("layers_6_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_6_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112194496))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116913152))))[name = string("layers_6_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_6_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116914752))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117111424))))[name = string("layers_6_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_7_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117111744))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118684672))))[name = string("layers_7_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_7_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118686784))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118883456))))[name = string("layers_7_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_7_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118883776))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119080448))))[name = string("layers_7_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_7_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119080768))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123799424))))[name = string("layers_7_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_7_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123805632))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128524288))))[name = string("layers_7_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_7_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128530496))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133249152))))[name = string("layers_7_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_7_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133250752))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133447424))))[name = string("layers_7_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_8_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133447744))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135020672))))[name = string("layers_8_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_8_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135022784))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135219456))))[name = string("layers_8_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_8_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135219776))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135416448))))[name = string("layers_8_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_8_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135416768))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140135424))))[name = string("layers_8_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_8_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140141632))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144860288))))[name = string("layers_8_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_8_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144866496))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149585152))))[name = string("layers_8_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_8_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149586752))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149783424))))[name = string("layers_8_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [4096, 1536, 1, 1]> layers_9_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4096, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149783744))), lut = tensor<fp16, [128, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152929536))))[name = string("layers_9_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [512, 1536, 1, 1]> layers_9_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [512, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152933696))), lut = tensor<fp16, [16, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153326976))))[name = string("layers_9_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [512, 1536, 1, 1]> layers_9_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [512, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153327552))), lut = tensor<fp16, [16, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153720832))))[name = string("layers_9_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_9_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153721408))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158440064))))[name = string("layers_9_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_9_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158446272))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163164928))))[name = string("layers_9_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_9_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163171136))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167889792))))[name = string("layers_9_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_9_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167891392))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168088064))))[name = string("layers_9_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_10_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168088384))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169661312))))[name = string("layers_10_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_10_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169663424))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169860096))))[name = string("layers_10_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_10_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169860416))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170057088))))[name = string("layers_10_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_10_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170057408))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174776064))))[name = string("layers_10_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_10_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174782272))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179500928))))[name = string("layers_10_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_10_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179507136))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184225792))))[name = string("layers_10_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_10_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184227392))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184424064))))[name = string("layers_10_per_layer_input_gate_weight_palettized")];
+            tensor<fp16, [2048, 1536, 1, 1]> layers_11_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184424384))), lut = tensor<fp16, [64, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185997312))))[name = string("layers_11_self_attn_q_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_11_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185999424))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186196096))))[name = string("layers_11_self_attn_k_proj_weight_palettized")];
+            tensor<fp16, [256, 1536, 1, 1]> layers_11_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186196416))), lut = tensor<fp16, [8, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186393088))))[name = string("layers_11_self_attn_v_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_11_mlp_gate_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186393408))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191112064))))[name = string("layers_11_mlp_gate_proj_weight_palettized")];
+            tensor<fp16, [6144, 1536, 1, 1]> layers_11_mlp_up_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [6144, 1536, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191118272))), lut = tensor<fp16, [192, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195836928))))[name = string("layers_11_mlp_up_proj_weight_palettized")];
+            tensor<fp16, [1536, 6144, 1, 1]> layers_11_mlp_down_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195843136))), lut = tensor<fp16, [48, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200561792))))[name = string("layers_11_mlp_down_proj_weight_palettized")];
+            tensor<fp16, [256, 1536]> layers_11_per_layer_input_gate_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200563392))), lut = tensor<fp16, [8, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200760064))))[name = string("layers_11_per_layer_input_gate_weight_palettized")];
+            int32 var_732_batch_dims_0 = const()[name = string("op_732_batch_dims_0"), val = int32(0)];
+            bool var_732_validate_indices_0 = const()[name = string("op_732_validate_indices_0"), val = bool(false)];
+            string position_ids_to_int16_dtype_0 = const()[name = string("position_ids_to_int16_dtype_0"), val = string("int16")];
+            string cast_126_dtype_0 = const()[name = string("cast_126_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]> position_ids_to_int16 = cast(dtype = position_ids_to_int16_dtype_0, x = position_ids)[name = string("cast_5")];
+            tensor<int32, [1]> cast_126 = cast(dtype = cast_126_dtype_0, x = position_ids_to_int16)[name = string("cast_4")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_126, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_24 = const()[name = string("slice_by_index_24"), val = int32(1024)];
+            tensor<int32, [1]> add_0 = add(x = cast_126, y = slice_by_index_24)[name = string("add_0")];
+            tensor<int32, [1]> select_0 = select(a = cast_126, 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 = const()[name = string("slice_by_index_0"), val = int32(1024)];
+            tensor<int32, [1]> add_0_1 = add(x = cast_0, y = slice_by_index_0)[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_732_cast_uint16_cast_uint16_axis_0 = const()[name = string("op_732_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 256]> op_732_cast_uint16_cast_uint16 = gather(axis = op_732_cast_uint16_cast_uint16_axis_0, batch_dims = var_732_batch_dims_0, indices = select_0_1, validate_indices = var_732_validate_indices_0, x = cos_sliding_palettized)[name = string("op_732_cast_uint16_cast_uint16")];
+            tensor<int32, [1]> var_734_axes_0 = const()[name = string("op_734_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 256]> var_734 = expand_dims(axes = var_734_axes_0, x = op_732_cast_uint16_cast_uint16)[name = string("op_734")];
+            tensor<int32, [1]> cos_1_axes_0 = const()[name = string("cos_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 1, 256]> cos_1 = expand_dims(axes = cos_1_axes_0, x = var_734)[name = string("cos_1")];
+            int32 var_737 = const()[name = string("op_737"), val = int32(0)];
+            int32 var_738_batch_dims_0 = const()[name = string("op_738_batch_dims_0"), val = int32(0)];
+            bool var_738_validate_indices_0 = const()[name = string("op_738_validate_indices_0"), val = bool(false)];
+            string position_ids_to_uint16_dtype_0 = const()[name = string("position_ids_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1]> position_ids_to_uint16 = cast(dtype = position_ids_to_uint16_dtype_0, x = position_ids)[name = string("cast_1")];
+            tensor<fp16, [1, 256]> var_738_cast_uint16 = gather(axis = var_737, batch_dims = var_738_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_738_validate_indices_0, x = sin_sliding_palettized)[name = string("op_738_cast_uint16")];
+            tensor<int32, [1]> var_740_axes_0 = const()[name = string("op_740_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 256]> var_740 = expand_dims(axes = var_740_axes_0, x = var_738_cast_uint16)[name = string("op_740")];
+            tensor<int32, [1]> sin_1_axes_0 = const()[name = string("sin_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 1, 256]> sin_1 = expand_dims(axes = sin_1_axes_0, x = var_740)[name = string("sin_1")];
+            int32 var_743 = const()[name = string("op_743"), val = int32(0)];
+            int32 var_744_batch_dims_0 = const()[name = string("op_744_batch_dims_0"), val = int32(0)];
+            bool var_744_validate_indices_0 = const()[name = string("op_744_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [1, 512]> var_744_cast_uint16 = gather(axis = var_743, batch_dims = var_744_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_744_validate_indices_0, x = cos_full_palettized)[name = string("op_744_cast_uint16")];
+            tensor<int32, [1]> var_746_axes_0 = const()[name = string("op_746_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512]> var_746 = expand_dims(axes = var_746_axes_0, x = var_744_cast_uint16)[name = string("op_746")];
+            tensor<int32, [1]> cos_axes_0 = const()[name = string("cos_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 1, 512]> cos = expand_dims(axes = cos_axes_0, x = var_746)[name = string("cos")];
+            int32 var_749 = const()[name = string("op_749"), val = int32(0)];
+            int32 var_750_batch_dims_0 = const()[name = string("op_750_batch_dims_0"), val = int32(0)];
+            bool var_750_validate_indices_0 = const()[name = string("op_750_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [1, 512]> var_750_cast_uint16 = gather(axis = var_749, batch_dims = var_750_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_750_validate_indices_0, x = sin_full_palettized)[name = string("op_750_cast_uint16")];
+            tensor<int32, [1]> var_752_axes_0 = const()[name = string("op_752_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512]> var_752 = expand_dims(axes = var_752_axes_0, x = var_750_cast_uint16)[name = string("op_752")];
+            tensor<int32, [1]> sin_axes_0 = const()[name = string("sin_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 1, 512]> sin = expand_dims(axes = sin_axes_0, x = var_752)[name = string("sin")];
+            int32 var_759 = const()[name = string("op_759"), 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_765_cast_fp16 = mul(x = hidden_states, y = const_0_promoted_to_fp16)[name = string("op_765_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_759, interleave = input_1_interleave_0, values = (hidden_states, var_765_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_757_to_fp16 = const()[name = string("op_757_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_757_to_fp16, x = input_1_cast_fp16)[name = string("normed_1_cast_fp16")];
+            tensor<int32, [2]> var_770_split_sizes_0 = const()[name = string("op_770_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_770_axis_0 = const()[name = string("op_770_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_770_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_770_cast_fp16_1 = split(axis = var_770_axis_0, split_sizes = var_770_split_sizes_0, x = normed_1_cast_fp16)[name = string("op_770_cast_fp16")];
+            tensor<fp16, [1536]> const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200760384)))];
+            tensor<fp16, [1, 1, 1536]> var_773_cast_fp16 = mul(x = var_770_cast_fp16_0, y = const_1_to_fp16)[name = string("op_773_cast_fp16")];
+            tensor<int32, [3]> var_778 = const()[name = string("op_778"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_781_axes_0 = const()[name = string("op_781_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_779 = transpose(perm = var_778, x = var_773_cast_fp16)[name = string("transpose_179")];
+            tensor<fp16, [1, 1536, 1, 1]> var_781 = expand_dims(axes = var_781_axes_0, x = var_779)[name = string("op_781")];
+            string var_797_pad_type_0 = const()[name = string("op_797_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_797_strides_0 = const()[name = string("op_797_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_797_pad_0 = const()[name = string("op_797_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_797_dilations_0 = const()[name = string("op_797_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_797_groups_0 = const()[name = string("op_797_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_797 = conv(dilations = var_797_dilations_0, groups = var_797_groups_0, pad = var_797_pad_0, pad_type = var_797_pad_type_0, strides = var_797_strides_0, weight = layers_0_self_attn_q_proj_weight_palettized, x = var_781)[name = string("op_797")];
+            tensor<int32, [4]> var_802 = const()[name = string("op_802"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_803 = reshape(shape = var_802, x = var_797)[name = string("op_803")];
+            tensor<int32, [4]> var_808 = const()[name = string("op_808"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_818 = const()[name = string("op_818"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_809 = transpose(perm = var_808, x = var_803)[name = string("transpose_178")];
+            tensor<fp16, [1, 8, 256]> x_3 = reshape(shape = var_818, x = var_809)[name = string("x_3")];
+            int32 var_824 = const()[name = string("op_824"), val = int32(-1)];
+            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_830_cast_fp16 = mul(x = x_3, y = const_2_promoted_to_fp16)[name = string("op_830_cast_fp16")];
+            bool input_5_interleave_0 = const()[name = string("input_5_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_5_cast_fp16 = concat(axis = var_824, interleave = input_5_interleave_0, values = (x_3, var_830_cast_fp16))[name = string("input_5_cast_fp16")];
+            tensor<int32, [1]> normed_5_axes_0 = const()[name = string("normed_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_822_to_fp16 = const()[name = string("op_822_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_5_cast_fp16 = layer_norm(axes = normed_5_axes_0, epsilon = var_822_to_fp16, x = input_5_cast_fp16)[name = string("normed_5_cast_fp16")];
+            tensor<int32, [2]> var_835_split_sizes_0 = const()[name = string("op_835_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_835_axis_0 = const()[name = string("op_835_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_835_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_835_cast_fp16_1 = split(axis = var_835_axis_0, split_sizes = var_835_split_sizes_0, x = normed_5_cast_fp16)[name = string("op_835_cast_fp16")];
+            tensor<fp16, [256]> const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200763520)))];
+            tensor<fp16, [1, 8, 256]> var_838_cast_fp16 = mul(x = var_835_cast_fp16_0, y = const_3_to_fp16)[name = string("op_838_cast_fp16")];
+            tensor<int32, [4]> var_844 = const()[name = string("op_844"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_3 = reshape(shape = var_844, x = var_838_cast_fp16)[name = string("q_3")];
+            tensor<fp16, [1, 8, 1, 256]> var_846 = mul(x = q_3, y = cos_1)[name = string("op_846")];
+            tensor<int32, [2]> var_847_split_sizes_0 = const()[name = string("op_847_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_847_axis_0 = const()[name = string("op_847_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_847_0, tensor<fp16, [1, 8, 1, 128]> var_847_1 = split(axis = var_847_axis_0, split_sizes = var_847_split_sizes_0, x = q_3)[name = string("op_847")];
+            fp16 const_4_promoted = const()[name = string("const_4_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_849 = mul(x = var_847_1, y = const_4_promoted)[name = string("op_849")];
+            int32 var_851 = const()[name = string("op_851"), val = int32(-1)];
+            bool var_852_interleave_0 = const()[name = string("op_852_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_852 = concat(axis = var_851, interleave = var_852_interleave_0, values = (var_849, var_847_0))[name = string("op_852")];
+            tensor<fp16, [1, 8, 1, 256]> var_853 = mul(x = var_852, y = sin_1)[name = string("op_853")];
+            tensor<fp16, [1, 8, 1, 256]> q_7 = add(x = var_846, y = var_853)[name = string("q_7")];
+            string var_866_pad_type_0 = const()[name = string("op_866_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_866_strides_0 = const()[name = string("op_866_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_866_pad_0 = const()[name = string("op_866_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_866_dilations_0 = const()[name = string("op_866_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_866_groups_0 = const()[name = string("op_866_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_866 = conv(dilations = var_866_dilations_0, groups = var_866_groups_0, pad = var_866_pad_0, pad_type = var_866_pad_type_0, strides = var_866_strides_0, weight = layers_0_self_attn_k_proj_weight_palettized, x = var_781)[name = string("op_866")];
+            tensor<int32, [4]> var_871 = const()[name = string("op_871"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_872 = reshape(shape = var_871, x = var_866)[name = string("op_872")];
+            tensor<int32, [4]> var_877 = const()[name = string("op_877"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_894_pad_type_0 = const()[name = string("op_894_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_894_strides_0 = const()[name = string("op_894_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_894_pad_0 = const()[name = string("op_894_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_894_dilations_0 = const()[name = string("op_894_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_894_groups_0 = const()[name = string("op_894_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_894 = conv(dilations = var_894_dilations_0, groups = var_894_groups_0, pad = var_894_pad_0, pad_type = var_894_pad_type_0, strides = var_894_strides_0, weight = layers_0_self_attn_v_proj_weight_palettized, x = var_781)[name = string("op_894")];
+            tensor<int32, [4]> var_899 = const()[name = string("op_899"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_900 = reshape(shape = var_899, x = var_894)[name = string("op_900")];
+            tensor<int32, [4]> var_905 = const()[name = string("op_905"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_915 = const()[name = string("op_915"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_878 = transpose(perm = var_877, x = var_872)[name = string("transpose_177")];
+            tensor<fp16, [1, 1, 256]> x_7 = reshape(shape = var_915, x = var_878)[name = string("x_7")];
+            int32 var_921 = const()[name = string("op_921"), val = int32(-1)];
+            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_927_cast_fp16 = mul(x = x_7, y = const_5_promoted_to_fp16)[name = string("op_927_cast_fp16")];
+            bool input_7_interleave_0 = const()[name = string("input_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_7_cast_fp16 = concat(axis = var_921, interleave = input_7_interleave_0, values = (x_7, var_927_cast_fp16))[name = string("input_7_cast_fp16")];
+            tensor<int32, [1]> normed_9_axes_0 = const()[name = string("normed_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_919_to_fp16 = const()[name = string("op_919_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_9_cast_fp16 = layer_norm(axes = normed_9_axes_0, epsilon = var_919_to_fp16, x = input_7_cast_fp16)[name = string("normed_9_cast_fp16")];
+            tensor<int32, [2]> var_932_split_sizes_0 = const()[name = string("op_932_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_932_axis_0 = const()[name = string("op_932_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_932_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_932_cast_fp16_1 = split(axis = var_932_axis_0, split_sizes = var_932_split_sizes_0, x = normed_9_cast_fp16)[name = string("op_932_cast_fp16")];
+            tensor<fp16, [256]> const_6_to_fp16 = const()[name = string("const_6_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200764096)))];
+            tensor<fp16, [1, 1, 256]> var_935_cast_fp16 = mul(x = var_932_cast_fp16_0, y = const_6_to_fp16)[name = string("op_935_cast_fp16")];
+            tensor<int32, [4]> var_941 = const()[name = string("op_941"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_5 = reshape(shape = var_941, x = var_935_cast_fp16)[name = string("q_5")];
+            fp16 var_948_promoted_to_fp16 = const()[name = string("op_948_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_906 = transpose(perm = var_905, x = var_900)[name = string("transpose_176")];
+            tensor<fp16, [1, 1, 1, 256]> var_949_cast_fp16 = pow(x = var_906, y = var_948_promoted_to_fp16)[name = string("op_949_cast_fp16")];
+            tensor<int32, [1]> var_954_axes_0 = const()[name = string("op_954_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_954_keep_dims_0 = const()[name = string("op_954_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_954_cast_fp16 = reduce_mean(axes = var_954_axes_0, keep_dims = var_954_keep_dims_0, x = var_949_cast_fp16)[name = string("op_954_cast_fp16")];
+            fp16 var_956_to_fp16 = const()[name = string("op_956_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_1_cast_fp16 = add(x = var_954_cast_fp16, y = var_956_to_fp16)[name = string("mean_sq_1_cast_fp16")];
+            fp16 var_963_to_fp16 = const()[name = string("op_963_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_964_cast_fp16 = pow(x = mean_sq_1_cast_fp16, y = var_963_to_fp16)[name = string("op_964_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_965_cast_fp16 = mul(x = var_906, y = var_964_cast_fp16)[name = string("op_965_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_971 = mul(x = q_5, y = cos_1)[name = string("op_971")];
+            tensor<int32, [2]> var_972_split_sizes_0 = const()[name = string("op_972_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_972_axis_0 = const()[name = string("op_972_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_972_0, tensor<fp16, [1, 1, 1, 128]> var_972_1 = split(axis = var_972_axis_0, split_sizes = var_972_split_sizes_0, x = q_5)[name = string("op_972")];
+            fp16 const_7_promoted = const()[name = string("const_7_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_974 = mul(x = var_972_1, y = const_7_promoted)[name = string("op_974")];
+            int32 var_976 = const()[name = string("op_976"), val = int32(-1)];
+            bool var_977_interleave_0 = const()[name = string("op_977_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_977 = concat(axis = var_976, interleave = var_977_interleave_0, values = (var_974, var_972_0))[name = string("op_977")];
+            tensor<fp16, [1, 1, 1, 256]> var_978 = mul(x = var_977, y = sin_1)[name = string("op_978")];
+            tensor<fp16, [1, 1, 1, 256]> input_9 = add(x = var_971, y = var_978)[name = string("input_9")];
+            tensor<fp16, [24, 1, 512, 512]> read_state_0 = read_state(input = kv_cache_0)[name = string("read_state_0")];
+            tensor<int32, [4]> var_983_begin_0 = const()[name = string("op_983_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_983_end_0 = const()[name = string("op_983_end_0"), val = tensor<int32, [4]>([1, 1, 512, 512])];
+            tensor<bool, [4]> var_983_end_mask_0 = const()[name = string("op_983_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_983_squeeze_mask_0 = const()[name = string("op_983_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_983_cast_fp16 = slice_by_index(begin = var_983_begin_0, end = var_983_end_0, end_mask = var_983_end_mask_0, squeeze_mask = var_983_squeeze_mask_0, x = read_state_0)[name = string("op_983_cast_fp16")];
+            tensor<int32, [1]> K_c_1_axes_0 = const()[name = string("K_c_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_1_cast_fp16 = expand_dims(axes = K_c_1_axes_0, x = var_983_cast_fp16)[name = string("K_c_1_cast_fp16")];
+            tensor<int32, [4]> var_988_begin_0 = const()[name = string("op_988_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> var_988_end_0 = const()[name = string("op_988_end_0"), val = tensor<int32, [4]>([13, 1, 512, 512])];
+            tensor<bool, [4]> var_988_end_mask_0 = const()[name = string("op_988_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_988_squeeze_mask_0 = const()[name = string("op_988_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_988_cast_fp16 = slice_by_index(begin = var_988_begin_0, end = var_988_end_0, end_mask = var_988_end_mask_0, squeeze_mask = var_988_squeeze_mask_0, x = read_state_0)[name = string("op_988_cast_fp16")];
+            tensor<int32, [1]> V_c_1_axes_0 = const()[name = string("V_c_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_1_cast_fp16 = expand_dims(axes = V_c_1_axes_0, x = var_988_cast_fp16)[name = string("V_c_1_cast_fp16")];
+            tensor<int32, [8]> kp_1_pad_0 = const()[name = string("kp_1_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_1_mode_0 = const()[name = string("kp_1_mode_0"), val = string("constant")];
+            fp16 const_8_to_fp16 = const()[name = string("const_8_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_1_cast_fp16 = pad(constant_val = const_8_to_fp16, mode = kp_1_mode_0, pad = kp_1_pad_0, x = input_9)[name = string("kp_1_cast_fp16")];
+            tensor<int32, [8]> vp_1_pad_0 = const()[name = string("vp_1_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_1_mode_0 = const()[name = string("vp_1_mode_0"), val = string("constant")];
+            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_1_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = vp_1_mode_0, pad = vp_1_pad_0, x = var_965_cast_fp16)[name = string("vp_1_cast_fp16")];
+            fp16 var_1003_promoted_to_fp16 = const()[name = string("op_1003_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1, 1, 512, 1]> var_1005_cast_fp16 = sub(x = var_1003_promoted_to_fp16, y = update_mask)[name = string("op_1005_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1006_cast_fp16 = mul(x = K_c_1_cast_fp16, y = var_1005_cast_fp16)[name = string("op_1006_cast_fp16")];
+            tensor<int32, [4]> var_1007_reps_0 = const()[name = string("op_1007_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_1007_cast_fp16 = tile(reps = var_1007_reps_0, x = kp_1_cast_fp16)[name = string("op_1007_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1008_cast_fp16 = mul(x = var_1007_cast_fp16, y = update_mask)[name = string("op_1008_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_1_cast_fp16 = add(x = var_1006_cast_fp16, y = var_1008_cast_fp16)[name = string("K_n_1_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1014_cast_fp16 = mul(x = V_c_1_cast_fp16, y = var_1005_cast_fp16)[name = string("op_1014_cast_fp16")];
+            tensor<int32, [4]> var_1015_reps_0 = const()[name = string("op_1015_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_1015_cast_fp16 = tile(reps = var_1015_reps_0, x = vp_1_cast_fp16)[name = string("op_1015_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1016_cast_fp16 = mul(x = var_1015_cast_fp16, y = update_mask)[name = string("op_1016_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_1_cast_fp16 = add(x = var_1014_cast_fp16, y = var_1016_cast_fp16)[name = string("V_n_1_cast_fp16")];
+            tensor<int32, [1]> var_1020_axes_0 = const()[name = string("op_1020_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_1020_cast_fp16 = squeeze(axes = var_1020_axes_0, x = K_n_1_cast_fp16)[name = string("op_1020_cast_fp16")];
+            tensor<int32, [4]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_1_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_1_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_1_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_0, begin_mask = kv_cache_0_internal_tensor_assign_1_begin_mask_0, end = concat_1, end_mask = kv_cache_0_internal_tensor_assign_1_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_1_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_1_stride_0, update = var_1020_cast_fp16, x = read_state_0)[name = string("kv_cache_0_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_1_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_24_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_24 = read_state(input = kv_cache_0)[name = string("coreml_update_state_24")];
+            tensor<int32, [1]> var_1027_axes_0 = const()[name = string("op_1027_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_1027_cast_fp16 = squeeze(axes = var_1027_axes_0, x = V_n_1_cast_fp16)[name = string("op_1027_cast_fp16")];
+            tensor<int32, [4]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_2_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_2_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_2_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_2, begin_mask = kv_cache_0_internal_tensor_assign_2_begin_mask_0, end = concat_3, end_mask = kv_cache_0_internal_tensor_assign_2_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_2_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_2_stride_0, update = var_1027_cast_fp16, x = coreml_update_state_24)[name = string("kv_cache_0_internal_tensor_assign_2_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_2_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_25_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_25 = read_state(input = kv_cache_0)[name = string("coreml_update_state_25")];
+            tensor<int32, [4]> var_1037_begin_0 = const()[name = string("op_1037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1037_end_0 = const()[name = string("op_1037_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_1037_end_mask_0 = const()[name = string("op_1037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_1037_cast_fp16 = slice_by_index(begin = var_1037_begin_0, end = var_1037_end_0, end_mask = var_1037_end_mask_0, x = K_n_1_cast_fp16)[name = string("op_1037_cast_fp16")];
+            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = var_1037_cast_fp16)[name = string("transpose_175")];
+            tensor<fp16, [8, 1, 512, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
+            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_0_cast_fp16 = reshape(shape = concat_4, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
+            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_174")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_1_cast_fp16 = reshape(shape = concat_5, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
+            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_1046_begin_0 = const()[name = string("op_1046_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1046_end_0 = const()[name = string("op_1046_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_1046_end_mask_0 = const()[name = string("op_1046_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_1046_cast_fp16 = slice_by_index(begin = var_1046_begin_0, end = var_1046_end_0, end_mask = var_1046_end_mask_0, x = V_n_1_cast_fp16)[name = string("op_1046_cast_fp16")];
+            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = var_1046_cast_fp16)[name = string("transpose_173")];
+            tensor<fp16, [8, 1, 512, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
+            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_2_cast_fp16 = reshape(shape = concat_6, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
+            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_172")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_3_cast_fp16 = reshape(shape = concat_7, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
+            tensor<int32, [4]> Ve_1_perm_0 = const()[name = string("Ve_1_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_1064_transpose_x_0 = const()[name = string("op_1064_transpose_x_0"), val = bool(false)];
+            bool var_1064_transpose_y_0 = const()[name = string("op_1064_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_171")];
+            tensor<fp16, [1, 8, 1, 512]> var_1064_cast_fp16 = matmul(transpose_x = var_1064_transpose_x_0, transpose_y = var_1064_transpose_y_0, x = q_7, y = transpose_48_cast_fp16)[name = string("op_1064_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_1071_cast_fp16 = add(x = var_1064_cast_fp16, y = causal_mask)[name = string("op_1071_cast_fp16")];
+            int32 var_1072 = const()[name = string("op_1072"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_1074_cast_fp16 = softmax(axis = var_1072, x = var_1071_cast_fp16)[name = string("op_1074_cast_fp16")];
+            bool var_1090_transpose_x_0 = const()[name = string("op_1090_transpose_x_0"), val = bool(false)];
+            bool var_1090_transpose_y_0 = const()[name = string("op_1090_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_1_cast_fp16 = transpose(perm = Ve_1_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_170")];
+            tensor<fp16, [1, 8, 1, 256]> var_1090_cast_fp16 = matmul(transpose_x = var_1090_transpose_x_0, transpose_y = var_1090_transpose_y_0, x = var_1074_cast_fp16, y = Ve_1_cast_fp16)[name = string("op_1090_cast_fp16")];
+            tensor<int32, [4]> var_1100 = const()[name = string("op_1100"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1107 = const()[name = string("op_1107"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_1101 = transpose(perm = var_1100, x = var_1090_cast_fp16)[name = string("transpose_169")];
+            tensor<fp16, [1, 1, 2048]> var_1108 = reshape(shape = var_1107, x = var_1101)[name = string("op_1108")];
+            tensor<int32, [3]> var_1112 = const()[name = string("op_1112"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_0_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200764672))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202337600))))[name = string("squeeze_0_palettized")];
+            string var_1128_pad_type_0 = const()[name = string("op_1128_pad_type_0"), val = string("valid")];
+            int32 var_1128_groups_0 = const()[name = string("op_1128_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_1128_strides_0 = const()[name = string("op_1128_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_1128_pad_0 = const()[name = string("op_1128_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_1128_dilations_0 = const()[name = string("op_1128_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_1113 = transpose(perm = var_1112, x = var_1108)[name = string("transpose_168")];
+            tensor<fp16, [1, 1536, 1]> var_1128 = conv(dilations = var_1128_dilations_0, groups = var_1128_groups_0, pad = var_1128_pad_0, pad_type = var_1128_pad_type_0, strides = var_1128_strides_0, weight = squeeze_0_palettized, x = var_1113)[name = string("op_1128")];
+            tensor<int32, [3]> var_1132 = const()[name = string("op_1132"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_1138 = const()[name = string("op_1138"), val = int32(-1)];
+            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_13 = transpose(perm = var_1132, x = var_1128)[name = string("transpose_167")];
+            tensor<fp16, [1, 1, 1536]> var_1144_cast_fp16 = mul(x = x_13, y = const_10_promoted_to_fp16)[name = string("op_1144_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_1138, interleave = input_15_interleave_0, values = (x_13, var_1144_cast_fp16))[name = string("input_15_cast_fp16")];
+            tensor<int32, [1]> normed_13_axes_0 = const()[name = string("normed_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1136_to_fp16 = const()[name = string("op_1136_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_1136_to_fp16, x = input_15_cast_fp16)[name = string("normed_13_cast_fp16")];
+            tensor<int32, [2]> var_1149_split_sizes_0 = const()[name = string("op_1149_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1149_axis_0 = const()[name = string("op_1149_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1149_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1149_cast_fp16_1 = split(axis = var_1149_axis_0, split_sizes = var_1149_split_sizes_0, x = normed_13_cast_fp16)[name = string("op_1149_cast_fp16")];
+            tensor<fp16, [1536]> const_11_to_fp16 = const()[name = string("const_11_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202339200)))];
+            tensor<fp16, [1, 1, 1536]> var_1152_cast_fp16 = mul(x = var_1149_cast_fp16_0, y = const_11_to_fp16)[name = string("op_1152_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_17_cast_fp16 = add(x = hidden_states, y = var_1152_cast_fp16)[name = string("x_17_cast_fp16")];
+            int32 var_1160 = const()[name = string("op_1160"), val = int32(-1)];
+            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1166_cast_fp16 = mul(x = x_17_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_1166_cast_fp16")];
+            bool input_17_interleave_0 = const()[name = string("input_17_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_17_cast_fp16 = concat(axis = var_1160, interleave = input_17_interleave_0, values = (x_17_cast_fp16, var_1166_cast_fp16))[name = string("input_17_cast_fp16")];
+            tensor<int32, [1]> normed_17_axes_0 = const()[name = string("normed_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1158_to_fp16 = const()[name = string("op_1158_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_1158_to_fp16, x = input_17_cast_fp16)[name = string("normed_17_cast_fp16")];
+            tensor<int32, [2]> var_1171_split_sizes_0 = const()[name = string("op_1171_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1171_axis_0 = const()[name = string("op_1171_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1171_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1171_cast_fp16_1 = split(axis = var_1171_axis_0, split_sizes = var_1171_split_sizes_0, x = normed_17_cast_fp16)[name = string("op_1171_cast_fp16")];
+            tensor<fp16, [1536]> const_13_to_fp16 = const()[name = string("const_13_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202342336)))];
+            tensor<fp16, [1, 1, 1536]> var_1174_cast_fp16 = mul(x = var_1171_cast_fp16_0, y = const_13_to_fp16)[name = string("op_1174_cast_fp16")];
+            tensor<int32, [3]> var_1184 = const()[name = string("op_1184"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_19_axes_0 = const()[name = string("input_19_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1185 = transpose(perm = var_1184, x = var_1174_cast_fp16)[name = string("transpose_166")];
+            tensor<fp16, [1, 1536, 1, 1]> input_19 = expand_dims(axes = input_19_axes_0, x = var_1185)[name = string("input_19")];
+            string var_1198_pad_type_0 = const()[name = string("op_1198_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1198_strides_0 = const()[name = string("op_1198_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1198_pad_0 = const()[name = string("op_1198_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1198_dilations_0 = const()[name = string("op_1198_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1198_groups_0 = const()[name = string("op_1198_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_1198 = conv(dilations = var_1198_dilations_0, groups = var_1198_groups_0, pad = var_1198_pad_0, pad_type = var_1198_pad_type_0, strides = var_1198_strides_0, weight = layers_0_mlp_gate_proj_weight_palettized, x = input_19)[name = string("op_1198")];
+            string var_1200_mode_0 = const()[name = string("op_1200_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_1200 = gelu(mode = var_1200_mode_0, x = var_1198)[name = string("op_1200")];
+            string var_1211_pad_type_0 = const()[name = string("op_1211_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1211_strides_0 = const()[name = string("op_1211_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1211_pad_0 = const()[name = string("op_1211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1211_dilations_0 = const()[name = string("op_1211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1211_groups_0 = const()[name = string("op_1211_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_1211 = conv(dilations = var_1211_dilations_0, groups = var_1211_groups_0, pad = var_1211_pad_0, pad_type = var_1211_pad_type_0, strides = var_1211_strides_0, weight = layers_0_mlp_up_proj_weight_palettized, x = input_19)[name = string("op_1211")];
+            tensor<fp16, [1, 6144, 1, 1]> input_21 = mul(x = var_1200, y = var_1211)[name = string("input_21")];
+            string var_1223_pad_type_0 = const()[name = string("op_1223_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1223_strides_0 = const()[name = string("op_1223_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1223_pad_0 = const()[name = string("op_1223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1223_dilations_0 = const()[name = string("op_1223_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1223_groups_0 = const()[name = string("op_1223_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_1223 = conv(dilations = var_1223_dilations_0, groups = var_1223_groups_0, pad = var_1223_pad_0, pad_type = var_1223_pad_type_0, strides = var_1223_strides_0, weight = layers_0_mlp_down_proj_weight_palettized, x = input_21)[name = string("op_1223")];
+            tensor<int32, [1]> var_1225_axes_0 = const()[name = string("op_1225_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1225 = squeeze(axes = var_1225_axes_0, x = var_1223)[name = string("op_1225")];
+            tensor<int32, [3]> var_1229 = const()[name = string("op_1229"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_1235 = const()[name = string("op_1235"), val = int32(-1)];
+            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_21 = transpose(perm = var_1229, x = var_1225)[name = string("transpose_165")];
+            tensor<fp16, [1, 1, 1536]> var_1241_cast_fp16 = mul(x = x_21, y = const_14_promoted_to_fp16)[name = string("op_1241_cast_fp16")];
+            bool input_23_interleave_0 = const()[name = string("input_23_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_23_cast_fp16 = concat(axis = var_1235, interleave = input_23_interleave_0, values = (x_21, var_1241_cast_fp16))[name = string("input_23_cast_fp16")];
+            tensor<int32, [1]> normed_21_axes_0 = const()[name = string("normed_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1233_to_fp16 = const()[name = string("op_1233_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_1233_to_fp16, x = input_23_cast_fp16)[name = string("normed_21_cast_fp16")];
+            tensor<int32, [2]> var_1246_split_sizes_0 = const()[name = string("op_1246_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1246_axis_0 = const()[name = string("op_1246_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1246_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1246_cast_fp16_1 = split(axis = var_1246_axis_0, split_sizes = var_1246_split_sizes_0, x = normed_21_cast_fp16)[name = string("op_1246_cast_fp16")];
+            tensor<fp16, [1536]> const_15_to_fp16 = const()[name = string("const_15_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202345472)))];
+            tensor<fp16, [1, 1, 1536]> var_1249_cast_fp16 = mul(x = var_1246_cast_fp16_0, y = const_15_to_fp16)[name = string("op_1249_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_13_cast_fp16 = add(x = x_17_cast_fp16, y = var_1249_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
+            tensor<fp16, [256]> linear_0_bias_0 = const()[name = string("linear_0_bias_0"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202348608)))];
+            tensor<fp16, [1, 1, 256]> var_1260 = linear(bias = linear_0_bias_0, weight = layers_0_per_layer_input_gate_weight_palettized, x = hidden_states_13_cast_fp16)[name = string("linear_0")];
+            string gated_1_mode_0 = const()[name = string("gated_1_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_1 = gelu(mode = gated_1_mode_0, x = var_1260)[name = string("gated_1")];
+            tensor<int32, [3]> var_1277_begin_0 = const()[name = string("op_1277_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1277_end_0 = const()[name = string("op_1277_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_1277_end_mask_0 = const()[name = string("op_1277_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_1277_cast_fp16 = slice_by_index(begin = var_1277_begin_0, end = var_1277_end_0, end_mask = var_1277_end_mask_0, x = per_layer_combined)[name = string("op_1277_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_27_cast_fp16 = mul(x = gated_1, y = var_1277_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_0_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202349184))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202545856))))[name = string("layers_0_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1536]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202547456)))];
+            tensor<fp16, [1, 1, 1536]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")];
+            int32 var_1286 = const()[name = string("op_1286"), val = int32(-1)];
+            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1292_cast_fp16 = mul(x = linear_1_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_1292_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_1286, interleave = input_29_interleave_0, values = (linear_1_cast_fp16, var_1292_cast_fp16))[name = string("input_29_cast_fp16")];
+            tensor<int32, [1]> normed_25_axes_0 = const()[name = string("normed_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1284_to_fp16 = const()[name = string("op_1284_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_1284_to_fp16, x = input_29_cast_fp16)[name = string("normed_25_cast_fp16")];
+            tensor<int32, [2]> var_1297_split_sizes_0 = const()[name = string("op_1297_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1297_axis_0 = const()[name = string("op_1297_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1297_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1297_cast_fp16_1 = split(axis = var_1297_axis_0, split_sizes = var_1297_split_sizes_0, x = normed_25_cast_fp16)[name = string("op_1297_cast_fp16")];
+            tensor<fp16, [1536]> const_17_to_fp16 = const()[name = string("const_17_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202550592)))];
+            tensor<fp16, [1, 1, 1536]> var_1300_cast_fp16 = mul(x = var_1297_cast_fp16_0, y = const_17_to_fp16)[name = string("op_1300_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_17 = add(x = hidden_states_13_cast_fp16, y = var_1300_cast_fp16)[name = string("hidden_states_17")];
+            tensor<fp16, [1]> layers_0_layer_scalar_to_fp16 = const()[name = string("layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.24p-6])];
+            tensor<fp16, [1, 1, 1536]> x_29_cast_fp16 = mul(x = hidden_states_17, y = layers_0_layer_scalar_to_fp16)[name = string("x_29_cast_fp16")];
+            int32 var_1308 = const()[name = string("op_1308"), 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_1314_cast_fp16 = mul(x = x_29_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_1314_cast_fp16")];
+            bool input_31_interleave_0 = const()[name = string("input_31_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_31_cast_fp16 = concat(axis = var_1308, interleave = input_31_interleave_0, values = (x_29_cast_fp16, var_1314_cast_fp16))[name = string("input_31_cast_fp16")];
+            tensor<int32, [1]> normed_29_axes_0 = const()[name = string("normed_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1306_to_fp16 = const()[name = string("op_1306_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_1306_to_fp16, x = input_31_cast_fp16)[name = string("normed_29_cast_fp16")];
+            tensor<int32, [2]> var_1319_split_sizes_0 = const()[name = string("op_1319_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1319_axis_0 = const()[name = string("op_1319_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1319_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1319_cast_fp16_1 = split(axis = var_1319_axis_0, split_sizes = var_1319_split_sizes_0, x = normed_29_cast_fp16)[name = string("op_1319_cast_fp16")];
+            tensor<fp16, [1536]> const_19_to_fp16 = const()[name = string("const_19_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202553728)))];
+            tensor<fp16, [1, 1, 1536]> var_1322_cast_fp16 = mul(x = var_1319_cast_fp16_0, y = const_19_to_fp16)[name = string("op_1322_cast_fp16")];
+            tensor<int32, [3]> var_1330 = const()[name = string("op_1330"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_1333_axes_0 = const()[name = string("op_1333_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1331_cast_fp16 = transpose(perm = var_1330, x = var_1322_cast_fp16)[name = string("transpose_164")];
+            tensor<fp16, [1, 1536, 1, 1]> var_1333_cast_fp16 = expand_dims(axes = var_1333_axes_0, x = var_1331_cast_fp16)[name = string("op_1333_cast_fp16")];
+            string var_1349_pad_type_0 = const()[name = string("op_1349_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1349_strides_0 = const()[name = string("op_1349_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1349_pad_0 = const()[name = string("op_1349_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1349_dilations_0 = const()[name = string("op_1349_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1349_groups_0 = const()[name = string("op_1349_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_1349 = conv(dilations = var_1349_dilations_0, groups = var_1349_groups_0, pad = var_1349_pad_0, pad_type = var_1349_pad_type_0, strides = var_1349_strides_0, weight = layers_1_self_attn_q_proj_weight_palettized, x = var_1333_cast_fp16)[name = string("op_1349")];
+            tensor<int32, [4]> var_1354 = const()[name = string("op_1354"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_1355 = reshape(shape = var_1354, x = var_1349)[name = string("op_1355")];
+            tensor<int32, [4]> var_1360 = const()[name = string("op_1360"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_1370 = const()[name = string("op_1370"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_1361 = transpose(perm = var_1360, x = var_1355)[name = string("transpose_163")];
+            tensor<fp16, [1, 8, 256]> x_33 = reshape(shape = var_1370, x = var_1361)[name = string("x_33")];
+            int32 var_1376 = const()[name = string("op_1376"), val = int32(-1)];
+            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_1382_cast_fp16 = mul(x = x_33, y = const_20_promoted_to_fp16)[name = string("op_1382_cast_fp16")];
+            bool input_35_interleave_0 = const()[name = string("input_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_35_cast_fp16 = concat(axis = var_1376, interleave = input_35_interleave_0, values = (x_33, var_1382_cast_fp16))[name = string("input_35_cast_fp16")];
+            tensor<int32, [1]> normed_33_axes_0 = const()[name = string("normed_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1374_to_fp16 = const()[name = string("op_1374_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_33_cast_fp16 = layer_norm(axes = normed_33_axes_0, epsilon = var_1374_to_fp16, x = input_35_cast_fp16)[name = string("normed_33_cast_fp16")];
+            tensor<int32, [2]> var_1387_split_sizes_0 = const()[name = string("op_1387_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_1387_axis_0 = const()[name = string("op_1387_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_1387_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_1387_cast_fp16_1 = split(axis = var_1387_axis_0, split_sizes = var_1387_split_sizes_0, x = normed_33_cast_fp16)[name = string("op_1387_cast_fp16")];
+            tensor<fp16, [256]> const_21_to_fp16 = const()[name = string("const_21_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202556864)))];
+            tensor<fp16, [1, 8, 256]> var_1390_cast_fp16 = mul(x = var_1387_cast_fp16_0, y = const_21_to_fp16)[name = string("op_1390_cast_fp16")];
+            tensor<int32, [4]> var_1396 = const()[name = string("op_1396"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_11 = reshape(shape = var_1396, x = var_1390_cast_fp16)[name = string("q_11")];
+            tensor<fp16, [1, 8, 1, 256]> var_1398 = mul(x = q_11, y = cos_1)[name = string("op_1398")];
+            tensor<int32, [2]> var_1399_split_sizes_0 = const()[name = string("op_1399_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_1399_axis_0 = const()[name = string("op_1399_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_1399_0, tensor<fp16, [1, 8, 1, 128]> var_1399_1 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = q_11)[name = string("op_1399")];
+            fp16 const_22_promoted = const()[name = string("const_22_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_1401 = mul(x = var_1399_1, y = const_22_promoted)[name = string("op_1401")];
+            int32 var_1403 = const()[name = string("op_1403"), val = int32(-1)];
+            bool var_1404_interleave_0 = const()[name = string("op_1404_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_1404 = concat(axis = var_1403, interleave = var_1404_interleave_0, values = (var_1401, var_1399_0))[name = string("op_1404")];
+            tensor<fp16, [1, 8, 1, 256]> var_1405 = mul(x = var_1404, y = sin_1)[name = string("op_1405")];
+            tensor<fp16, [1, 8, 1, 256]> q_15 = add(x = var_1398, y = var_1405)[name = string("q_15")];
+            string var_1418_pad_type_0 = const()[name = string("op_1418_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1418_strides_0 = const()[name = string("op_1418_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1418_pad_0 = const()[name = string("op_1418_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1418_dilations_0 = const()[name = string("op_1418_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1418_groups_0 = const()[name = string("op_1418_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1418 = conv(dilations = var_1418_dilations_0, groups = var_1418_groups_0, pad = var_1418_pad_0, pad_type = var_1418_pad_type_0, strides = var_1418_strides_0, weight = layers_1_self_attn_k_proj_weight_palettized, x = var_1333_cast_fp16)[name = string("op_1418")];
+            tensor<int32, [4]> var_1423 = const()[name = string("op_1423"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_1424 = reshape(shape = var_1423, x = var_1418)[name = string("op_1424")];
+            tensor<int32, [4]> var_1429 = const()[name = string("op_1429"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_1446_pad_type_0 = const()[name = string("op_1446_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1446_strides_0 = const()[name = string("op_1446_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1446_pad_0 = const()[name = string("op_1446_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1446_dilations_0 = const()[name = string("op_1446_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1446_groups_0 = const()[name = string("op_1446_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1446 = conv(dilations = var_1446_dilations_0, groups = var_1446_groups_0, pad = var_1446_pad_0, pad_type = var_1446_pad_type_0, strides = var_1446_strides_0, weight = layers_1_self_attn_v_proj_weight_palettized, x = var_1333_cast_fp16)[name = string("op_1446")];
+            tensor<int32, [4]> var_1451 = const()[name = string("op_1451"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_1452 = reshape(shape = var_1451, x = var_1446)[name = string("op_1452")];
+            tensor<int32, [4]> var_1457 = const()[name = string("op_1457"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_1467 = const()[name = string("op_1467"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_1430 = transpose(perm = var_1429, x = var_1424)[name = string("transpose_162")];
+            tensor<fp16, [1, 1, 256]> x_37 = reshape(shape = var_1467, x = var_1430)[name = string("x_37")];
+            int32 var_1473 = const()[name = string("op_1473"), val = int32(-1)];
+            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_1479_cast_fp16 = mul(x = x_37, y = const_23_promoted_to_fp16)[name = string("op_1479_cast_fp16")];
+            bool input_37_interleave_0 = const()[name = string("input_37_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_37_cast_fp16 = concat(axis = var_1473, interleave = input_37_interleave_0, values = (x_37, var_1479_cast_fp16))[name = string("input_37_cast_fp16")];
+            tensor<int32, [1]> normed_37_axes_0 = const()[name = string("normed_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1471_to_fp16 = const()[name = string("op_1471_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_37_cast_fp16 = layer_norm(axes = normed_37_axes_0, epsilon = var_1471_to_fp16, x = input_37_cast_fp16)[name = string("normed_37_cast_fp16")];
+            tensor<int32, [2]> var_1484_split_sizes_0 = const()[name = string("op_1484_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_1484_axis_0 = const()[name = string("op_1484_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_1484_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_1484_cast_fp16_1 = split(axis = var_1484_axis_0, split_sizes = var_1484_split_sizes_0, x = normed_37_cast_fp16)[name = string("op_1484_cast_fp16")];
+            tensor<fp16, [256]> const_24_to_fp16 = const()[name = string("const_24_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202557440)))];
+            tensor<fp16, [1, 1, 256]> var_1487_cast_fp16 = mul(x = var_1484_cast_fp16_0, y = const_24_to_fp16)[name = string("op_1487_cast_fp16")];
+            tensor<int32, [4]> var_1493 = const()[name = string("op_1493"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_13 = reshape(shape = var_1493, x = var_1487_cast_fp16)[name = string("q_13")];
+            fp16 var_1500_promoted_to_fp16 = const()[name = string("op_1500_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_1458 = transpose(perm = var_1457, x = var_1452)[name = string("transpose_161")];
+            tensor<fp16, [1, 1, 1, 256]> var_1501_cast_fp16 = pow(x = var_1458, y = var_1500_promoted_to_fp16)[name = string("op_1501_cast_fp16")];
+            tensor<int32, [1]> var_1506_axes_0 = const()[name = string("op_1506_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_1506_keep_dims_0 = const()[name = string("op_1506_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_1506_cast_fp16 = reduce_mean(axes = var_1506_axes_0, keep_dims = var_1506_keep_dims_0, x = var_1501_cast_fp16)[name = string("op_1506_cast_fp16")];
+            fp16 var_1508_to_fp16 = const()[name = string("op_1508_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_3_cast_fp16 = add(x = var_1506_cast_fp16, y = var_1508_to_fp16)[name = string("mean_sq_3_cast_fp16")];
+            fp16 var_1515_to_fp16 = const()[name = string("op_1515_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_1516_cast_fp16 = pow(x = mean_sq_3_cast_fp16, y = var_1515_to_fp16)[name = string("op_1516_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_1517_cast_fp16 = mul(x = var_1458, y = var_1516_cast_fp16)[name = string("op_1517_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_1523 = mul(x = q_13, y = cos_1)[name = string("op_1523")];
+            tensor<int32, [2]> var_1524_split_sizes_0 = const()[name = string("op_1524_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_1524_axis_0 = const()[name = string("op_1524_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_1524_0, tensor<fp16, [1, 1, 1, 128]> var_1524_1 = split(axis = var_1524_axis_0, split_sizes = var_1524_split_sizes_0, x = q_13)[name = string("op_1524")];
+            fp16 const_25_promoted = const()[name = string("const_25_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_1526 = mul(x = var_1524_1, y = const_25_promoted)[name = string("op_1526")];
+            int32 var_1528 = const()[name = string("op_1528"), val = int32(-1)];
+            bool var_1529_interleave_0 = const()[name = string("op_1529_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_1529 = concat(axis = var_1528, interleave = var_1529_interleave_0, values = (var_1526, var_1524_0))[name = string("op_1529")];
+            tensor<fp16, [1, 1, 1, 256]> var_1530 = mul(x = var_1529, y = sin_1)[name = string("op_1530")];
+            tensor<fp16, [1, 1, 1, 256]> input_39 = add(x = var_1523, y = var_1530)[name = string("input_39")];
+            tensor<int32, [4]> var_1535_begin_0 = const()[name = string("op_1535_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> var_1535_end_0 = const()[name = string("op_1535_end_0"), val = tensor<int32, [4]>([2, 1, 512, 512])];
+            tensor<bool, [4]> var_1535_end_mask_0 = const()[name = string("op_1535_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_1535_squeeze_mask_0 = const()[name = string("op_1535_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_1535_cast_fp16 = slice_by_index(begin = var_1535_begin_0, end = var_1535_end_0, end_mask = var_1535_end_mask_0, squeeze_mask = var_1535_squeeze_mask_0, x = coreml_update_state_25)[name = string("op_1535_cast_fp16")];
+            tensor<int32, [1]> K_c_3_axes_0 = const()[name = string("K_c_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_3_cast_fp16 = expand_dims(axes = K_c_3_axes_0, x = var_1535_cast_fp16)[name = string("K_c_3_cast_fp16")];
+            tensor<int32, [4]> var_1540_begin_0 = const()[name = string("op_1540_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> var_1540_end_0 = const()[name = string("op_1540_end_0"), val = tensor<int32, [4]>([14, 1, 512, 512])];
+            tensor<bool, [4]> var_1540_end_mask_0 = const()[name = string("op_1540_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_1540_squeeze_mask_0 = const()[name = string("op_1540_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_1540_cast_fp16 = slice_by_index(begin = var_1540_begin_0, end = var_1540_end_0, end_mask = var_1540_end_mask_0, squeeze_mask = var_1540_squeeze_mask_0, x = coreml_update_state_25)[name = string("op_1540_cast_fp16")];
+            tensor<int32, [1]> V_c_3_axes_0 = const()[name = string("V_c_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_3_cast_fp16 = expand_dims(axes = V_c_3_axes_0, x = var_1540_cast_fp16)[name = string("V_c_3_cast_fp16")];
+            tensor<int32, [8]> kp_3_pad_0 = const()[name = string("kp_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_3_mode_0 = const()[name = string("kp_3_mode_0"), val = string("constant")];
+            fp16 const_26_to_fp16 = const()[name = string("const_26_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_3_cast_fp16 = pad(constant_val = const_26_to_fp16, mode = kp_3_mode_0, pad = kp_3_pad_0, x = input_39)[name = string("kp_3_cast_fp16")];
+            tensor<int32, [8]> vp_3_pad_0 = const()[name = string("vp_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_3_mode_0 = const()[name = string("vp_3_mode_0"), val = string("constant")];
+            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_3_cast_fp16 = pad(constant_val = const_27_to_fp16, mode = vp_3_mode_0, pad = vp_3_pad_0, x = var_1517_cast_fp16)[name = string("vp_3_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1558_cast_fp16 = mul(x = K_c_3_cast_fp16, y = var_1005_cast_fp16)[name = string("op_1558_cast_fp16")];
+            tensor<int32, [4]> var_1559_reps_0 = const()[name = string("op_1559_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_1559_cast_fp16 = tile(reps = var_1559_reps_0, x = kp_3_cast_fp16)[name = string("op_1559_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1560_cast_fp16 = mul(x = var_1559_cast_fp16, y = update_mask)[name = string("op_1560_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_3_cast_fp16 = add(x = var_1558_cast_fp16, y = var_1560_cast_fp16)[name = string("K_n_3_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1566_cast_fp16 = mul(x = V_c_3_cast_fp16, y = var_1005_cast_fp16)[name = string("op_1566_cast_fp16")];
+            tensor<int32, [4]> var_1567_reps_0 = const()[name = string("op_1567_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_1567_cast_fp16 = tile(reps = var_1567_reps_0, x = vp_3_cast_fp16)[name = string("op_1567_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_1568_cast_fp16 = mul(x = var_1567_cast_fp16, y = update_mask)[name = string("op_1568_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_3_cast_fp16 = add(x = var_1566_cast_fp16, y = var_1568_cast_fp16)[name = string("V_n_3_cast_fp16")];
+            tensor<int32, [1]> var_1572_axes_0 = const()[name = string("op_1572_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_1572_cast_fp16 = squeeze(axes = var_1572_axes_0, x = K_n_3_cast_fp16)[name = string("op_1572_cast_fp16")];
+            tensor<int32, [4]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_3_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_3_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_3_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_8, begin_mask = kv_cache_0_internal_tensor_assign_3_begin_mask_0, end = concat_9, end_mask = kv_cache_0_internal_tensor_assign_3_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_3_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_3_stride_0, update = var_1572_cast_fp16, x = coreml_update_state_25)[name = string("kv_cache_0_internal_tensor_assign_3_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_3_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_26_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_26 = read_state(input = kv_cache_0)[name = string("coreml_update_state_26")];
+            tensor<int32, [1]> var_1579_axes_0 = const()[name = string("op_1579_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_1579_cast_fp16 = squeeze(axes = var_1579_axes_0, x = V_n_3_cast_fp16)[name = string("op_1579_cast_fp16")];
+            tensor<int32, [4]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_4_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_4_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_4_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_10, begin_mask = kv_cache_0_internal_tensor_assign_4_begin_mask_0, end = concat_11, end_mask = kv_cache_0_internal_tensor_assign_4_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_4_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_4_stride_0, update = var_1579_cast_fp16, x = coreml_update_state_26)[name = string("kv_cache_0_internal_tensor_assign_4_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_4_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_27_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_27 = read_state(input = kv_cache_0)[name = string("coreml_update_state_27")];
+            tensor<int32, [4]> var_1589_begin_0 = const()[name = string("op_1589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1589_end_0 = const()[name = string("op_1589_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_1589_end_mask_0 = const()[name = string("op_1589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_1589_cast_fp16 = slice_by_index(begin = var_1589_begin_0, end = var_1589_end_0, end_mask = var_1589_end_mask_0, x = K_n_3_cast_fp16)[name = string("op_1589_cast_fp16")];
+            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = var_1589_cast_fp16)[name = string("transpose_160")];
+            tensor<fp16, [8, 1, 512, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
+            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_4_cast_fp16 = reshape(shape = concat_12, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
+            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_159")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_5_cast_fp16 = reshape(shape = concat_13, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
+            tensor<int32, [4]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_1598_begin_0 = const()[name = string("op_1598_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1598_end_0 = const()[name = string("op_1598_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_1598_end_mask_0 = const()[name = string("op_1598_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_1598_cast_fp16 = slice_by_index(begin = var_1598_begin_0, end = var_1598_end_0, end_mask = var_1598_end_mask_0, x = V_n_3_cast_fp16)[name = string("op_1598_cast_fp16")];
+            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = var_1598_cast_fp16)[name = string("transpose_158")];
+            tensor<fp16, [8, 1, 512, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
+            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_6_cast_fp16 = reshape(shape = concat_14, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
+            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_157")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_7_cast_fp16 = reshape(shape = concat_15, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
+            tensor<int32, [4]> Ve_3_perm_0 = const()[name = string("Ve_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_1616_transpose_x_0 = const()[name = string("op_1616_transpose_x_0"), val = bool(false)];
+            bool var_1616_transpose_y_0 = const()[name = string("op_1616_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_156")];
+            tensor<fp16, [1, 8, 1, 512]> var_1616_cast_fp16 = matmul(transpose_x = var_1616_transpose_x_0, transpose_y = var_1616_transpose_y_0, x = q_15, y = transpose_49_cast_fp16)[name = string("op_1616_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_1623_cast_fp16 = add(x = var_1616_cast_fp16, y = causal_mask)[name = string("op_1623_cast_fp16")];
+            int32 var_1624 = const()[name = string("op_1624"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_1626_cast_fp16 = softmax(axis = var_1624, x = var_1623_cast_fp16)[name = string("op_1626_cast_fp16")];
+            bool var_1642_transpose_x_0 = const()[name = string("op_1642_transpose_x_0"), val = bool(false)];
+            bool var_1642_transpose_y_0 = const()[name = string("op_1642_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_3_cast_fp16 = transpose(perm = Ve_3_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_155")];
+            tensor<fp16, [1, 8, 1, 256]> var_1642_cast_fp16 = matmul(transpose_x = var_1642_transpose_x_0, transpose_y = var_1642_transpose_y_0, x = var_1626_cast_fp16, y = Ve_3_cast_fp16)[name = string("op_1642_cast_fp16")];
+            tensor<int32, [4]> var_1652 = const()[name = string("op_1652"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_1653 = transpose(perm = var_1652, x = var_1642_cast_fp16)[name = string("transpose_154")];
+            tensor<fp16, [1, 1, 2048]> var_1660 = reshape(shape = var_1659, x = var_1653)[name = string("op_1660")];
+            tensor<int32, [3]> var_1664 = const()[name = string("op_1664"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_1_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202558016))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204130944))))[name = string("squeeze_1_palettized")];
+            string var_1680_pad_type_0 = const()[name = string("op_1680_pad_type_0"), val = string("valid")];
+            int32 var_1680_groups_0 = const()[name = string("op_1680_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_1680_strides_0 = const()[name = string("op_1680_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_1680_pad_0 = const()[name = string("op_1680_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_1680_dilations_0 = const()[name = string("op_1680_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_1665 = transpose(perm = var_1664, x = var_1660)[name = string("transpose_153")];
+            tensor<fp16, [1, 1536, 1]> var_1680 = conv(dilations = var_1680_dilations_0, groups = var_1680_groups_0, pad = var_1680_pad_0, pad_type = var_1680_pad_type_0, strides = var_1680_strides_0, weight = squeeze_1_palettized, x = var_1665)[name = string("op_1680")];
+            tensor<int32, [3]> var_1684 = const()[name = string("op_1684"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_1690 = const()[name = string("op_1690"), val = int32(-1)];
+            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_43 = transpose(perm = var_1684, x = var_1680)[name = string("transpose_152")];
+            tensor<fp16, [1, 1, 1536]> var_1696_cast_fp16 = mul(x = x_43, y = const_28_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
+            bool input_45_interleave_0 = const()[name = string("input_45_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_45_cast_fp16 = concat(axis = var_1690, interleave = input_45_interleave_0, values = (x_43, var_1696_cast_fp16))[name = string("input_45_cast_fp16")];
+            tensor<int32, [1]> normed_41_axes_0 = const()[name = string("normed_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1688_to_fp16 = const()[name = string("op_1688_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_1688_to_fp16, x = input_45_cast_fp16)[name = string("normed_41_cast_fp16")];
+            tensor<int32, [2]> var_1701_split_sizes_0 = const()[name = string("op_1701_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1701_axis_0 = const()[name = string("op_1701_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1701_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1701_cast_fp16_1 = split(axis = var_1701_axis_0, split_sizes = var_1701_split_sizes_0, x = normed_41_cast_fp16)[name = string("op_1701_cast_fp16")];
+            tensor<fp16, [1536]> const_29_to_fp16 = const()[name = string("const_29_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204132544)))];
+            tensor<fp16, [1, 1, 1536]> var_1704_cast_fp16 = mul(x = var_1701_cast_fp16_0, y = const_29_to_fp16)[name = string("op_1704_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_47_cast_fp16 = add(x = x_29_cast_fp16, y = var_1704_cast_fp16)[name = string("x_47_cast_fp16")];
+            int32 var_1711 = const()[name = string("op_1711"), val = int32(-1)];
+            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1717_cast_fp16 = mul(x = x_47_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1717_cast_fp16")];
+            bool input_47_interleave_0 = const()[name = string("input_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_47_cast_fp16 = concat(axis = var_1711, interleave = input_47_interleave_0, values = (x_47_cast_fp16, var_1717_cast_fp16))[name = string("input_47_cast_fp16")];
+            tensor<int32, [1]> normed_45_axes_0 = const()[name = string("normed_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1709_to_fp16 = const()[name = string("op_1709_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_1709_to_fp16, x = input_47_cast_fp16)[name = string("normed_45_cast_fp16")];
+            tensor<int32, [2]> var_1722_split_sizes_0 = const()[name = string("op_1722_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1722_axis_0 = const()[name = string("op_1722_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1722_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1722_cast_fp16_1 = split(axis = var_1722_axis_0, split_sizes = var_1722_split_sizes_0, x = normed_45_cast_fp16)[name = string("op_1722_cast_fp16")];
+            tensor<fp16, [1536]> const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204135680)))];
+            tensor<fp16, [1, 1, 1536]> var_1725_cast_fp16 = mul(x = var_1722_cast_fp16_0, y = const_31_to_fp16)[name = string("op_1725_cast_fp16")];
+            tensor<int32, [3]> var_1738 = const()[name = string("op_1738"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_49_axes_0 = const()[name = string("input_49_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1739 = transpose(perm = var_1738, x = var_1725_cast_fp16)[name = string("transpose_151")];
+            tensor<fp16, [1, 1536, 1, 1]> input_49 = expand_dims(axes = input_49_axes_0, x = var_1739)[name = string("input_49")];
+            string var_1752_pad_type_0 = const()[name = string("op_1752_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1752_strides_0 = const()[name = string("op_1752_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1752_pad_0 = const()[name = string("op_1752_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1752_dilations_0 = const()[name = string("op_1752_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1752_groups_0 = const()[name = string("op_1752_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_1752 = conv(dilations = var_1752_dilations_0, groups = var_1752_groups_0, pad = var_1752_pad_0, pad_type = var_1752_pad_type_0, strides = var_1752_strides_0, weight = layers_1_mlp_gate_proj_weight_palettized, x = input_49)[name = string("op_1752")];
+            string var_1754_mode_0 = const()[name = string("op_1754_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_1754 = gelu(mode = var_1754_mode_0, x = var_1752)[name = string("op_1754")];
+            string var_1765_pad_type_0 = const()[name = string("op_1765_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1765_strides_0 = const()[name = string("op_1765_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1765_pad_0 = const()[name = string("op_1765_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1765_dilations_0 = const()[name = string("op_1765_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1765_groups_0 = const()[name = string("op_1765_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_1765 = conv(dilations = var_1765_dilations_0, groups = var_1765_groups_0, pad = var_1765_pad_0, pad_type = var_1765_pad_type_0, strides = var_1765_strides_0, weight = layers_1_mlp_up_proj_weight_palettized, x = input_49)[name = string("op_1765")];
+            tensor<fp16, [1, 6144, 1, 1]> input_51 = mul(x = var_1754, y = var_1765)[name = string("input_51")];
+            string var_1777_pad_type_0 = const()[name = string("op_1777_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1777_strides_0 = const()[name = string("op_1777_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1777_pad_0 = const()[name = string("op_1777_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1777_dilations_0 = const()[name = string("op_1777_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1777_groups_0 = const()[name = string("op_1777_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_1777 = conv(dilations = var_1777_dilations_0, groups = var_1777_groups_0, pad = var_1777_pad_0, pad_type = var_1777_pad_type_0, strides = var_1777_strides_0, weight = layers_1_mlp_down_proj_weight_palettized, x = input_51)[name = string("op_1777")];
+            tensor<int32, [1]> var_1779_axes_0 = const()[name = string("op_1779_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1779 = squeeze(axes = var_1779_axes_0, x = var_1777)[name = string("op_1779")];
+            tensor<int32, [3]> var_1783 = const()[name = string("op_1783"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_1789 = const()[name = string("op_1789"), val = int32(-1)];
+            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_51 = transpose(perm = var_1783, x = var_1779)[name = string("transpose_150")];
+            tensor<fp16, [1, 1, 1536]> var_1795_cast_fp16 = mul(x = x_51, y = const_32_promoted_to_fp16)[name = string("op_1795_cast_fp16")];
+            bool input_53_interleave_0 = const()[name = string("input_53_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_53_cast_fp16 = concat(axis = var_1789, interleave = input_53_interleave_0, values = (x_51, var_1795_cast_fp16))[name = string("input_53_cast_fp16")];
+            tensor<int32, [1]> normed_49_axes_0 = const()[name = string("normed_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1787_to_fp16 = const()[name = string("op_1787_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_1787_to_fp16, x = input_53_cast_fp16)[name = string("normed_49_cast_fp16")];
+            tensor<int32, [2]> var_1800_split_sizes_0 = const()[name = string("op_1800_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1800_axis_0 = const()[name = string("op_1800_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1800_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1800_cast_fp16_1 = split(axis = var_1800_axis_0, split_sizes = var_1800_split_sizes_0, x = normed_49_cast_fp16)[name = string("op_1800_cast_fp16")];
+            tensor<fp16, [1536]> const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204138816)))];
+            tensor<fp16, [1, 1, 1536]> var_1803_cast_fp16 = mul(x = var_1800_cast_fp16_0, y = const_33_to_fp16)[name = string("op_1803_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_27_cast_fp16 = add(x = x_47_cast_fp16, y = var_1803_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_1814 = linear(bias = linear_0_bias_0, weight = layers_1_per_layer_input_gate_weight_palettized, x = hidden_states_27_cast_fp16)[name = string("linear_2")];
+            string gated_3_mode_0 = const()[name = string("gated_3_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_3 = gelu(mode = gated_3_mode_0, x = var_1814)[name = string("gated_3")];
+            tensor<int32, [3]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
+            tensor<int32, [3]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = per_layer_combined)[name = string("op_1831_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_57_cast_fp16 = mul(x = gated_3, y = var_1831_cast_fp16)[name = string("input_57_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_1_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204141952))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204338624))))[name = string("layers_1_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_57_cast_fp16)[name = string("linear_3_cast_fp16")];
+            int32 var_1840 = const()[name = string("op_1840"), val = int32(-1)];
+            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1846_cast_fp16 = mul(x = linear_3_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_1846_cast_fp16")];
+            bool input_59_interleave_0 = const()[name = string("input_59_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_59_cast_fp16 = concat(axis = var_1840, interleave = input_59_interleave_0, values = (linear_3_cast_fp16, var_1846_cast_fp16))[name = string("input_59_cast_fp16")];
+            tensor<int32, [1]> normed_53_axes_0 = const()[name = string("normed_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1838_to_fp16 = const()[name = string("op_1838_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_1838_to_fp16, x = input_59_cast_fp16)[name = string("normed_53_cast_fp16")];
+            tensor<int32, [2]> var_1851_split_sizes_0 = const()[name = string("op_1851_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1851_axis_0 = const()[name = string("op_1851_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1851_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1851_cast_fp16_1 = split(axis = var_1851_axis_0, split_sizes = var_1851_split_sizes_0, x = normed_53_cast_fp16)[name = string("op_1851_cast_fp16")];
+            tensor<fp16, [1536]> const_35_to_fp16 = const()[name = string("const_35_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204340224)))];
+            tensor<fp16, [1, 1, 1536]> var_1854_cast_fp16 = mul(x = var_1851_cast_fp16_0, y = const_35_to_fp16)[name = string("op_1854_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_31_cast_fp16 = add(x = hidden_states_27_cast_fp16, y = var_1854_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
+            tensor<fp16, [1]> layers_1_layer_scalar_to_fp16 = const()[name = string("layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.c8p-3])];
+            tensor<fp16, [1, 1, 1536]> x_59_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = layers_1_layer_scalar_to_fp16)[name = string("x_59_cast_fp16")];
+            int32 var_1862 = const()[name = string("op_1862"), val = int32(-1)];
+            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_1868_cast_fp16 = mul(x = x_59_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_1868_cast_fp16")];
+            bool input_61_interleave_0 = const()[name = string("input_61_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_61_cast_fp16 = concat(axis = var_1862, interleave = input_61_interleave_0, values = (x_59_cast_fp16, var_1868_cast_fp16))[name = string("input_61_cast_fp16")];
+            tensor<int32, [1]> normed_57_axes_0 = const()[name = string("normed_57_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1860_to_fp16 = const()[name = string("op_1860_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_1860_to_fp16, x = input_61_cast_fp16)[name = string("normed_57_cast_fp16")];
+            tensor<int32, [2]> var_1873_split_sizes_0 = const()[name = string("op_1873_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_1873_axis_0 = const()[name = string("op_1873_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_1873_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_1873_cast_fp16_1 = split(axis = var_1873_axis_0, split_sizes = var_1873_split_sizes_0, x = normed_57_cast_fp16)[name = string("op_1873_cast_fp16")];
+            tensor<fp16, [1536]> const_37_to_fp16 = const()[name = string("const_37_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204343360)))];
+            tensor<fp16, [1, 1, 1536]> var_1876_cast_fp16 = mul(x = var_1873_cast_fp16_0, y = const_37_to_fp16)[name = string("op_1876_cast_fp16")];
+            tensor<int32, [3]> var_1884 = const()[name = string("op_1884"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_1887_axes_0 = const()[name = string("op_1887_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_1885_cast_fp16 = transpose(perm = var_1884, x = var_1876_cast_fp16)[name = string("transpose_149")];
+            tensor<fp16, [1, 1536, 1, 1]> var_1887_cast_fp16 = expand_dims(axes = var_1887_axes_0, x = var_1885_cast_fp16)[name = string("op_1887_cast_fp16")];
+            string var_1903_pad_type_0 = const()[name = string("op_1903_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1903_strides_0 = const()[name = string("op_1903_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1903_pad_0 = const()[name = string("op_1903_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1903_dilations_0 = const()[name = string("op_1903_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1903_groups_0 = const()[name = string("op_1903_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_1903 = conv(dilations = var_1903_dilations_0, groups = var_1903_groups_0, pad = var_1903_pad_0, pad_type = var_1903_pad_type_0, strides = var_1903_strides_0, weight = layers_2_self_attn_q_proj_weight_palettized, x = var_1887_cast_fp16)[name = string("op_1903")];
+            tensor<int32, [4]> var_1908 = const()[name = string("op_1908"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_1909 = reshape(shape = var_1908, x = var_1903)[name = string("op_1909")];
+            tensor<int32, [4]> var_1914 = const()[name = string("op_1914"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_1924 = const()[name = string("op_1924"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_1915 = transpose(perm = var_1914, x = var_1909)[name = string("transpose_148")];
+            tensor<fp16, [1, 8, 256]> x_63 = reshape(shape = var_1924, x = var_1915)[name = string("x_63")];
+            int32 var_1930 = const()[name = string("op_1930"), val = int32(-1)];
+            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_1936_cast_fp16 = mul(x = x_63, y = const_38_promoted_to_fp16)[name = string("op_1936_cast_fp16")];
+            bool input_65_interleave_0 = const()[name = string("input_65_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_65_cast_fp16 = concat(axis = var_1930, interleave = input_65_interleave_0, values = (x_63, var_1936_cast_fp16))[name = string("input_65_cast_fp16")];
+            tensor<int32, [1]> normed_61_axes_0 = const()[name = string("normed_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_1928_to_fp16 = const()[name = string("op_1928_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_61_cast_fp16 = layer_norm(axes = normed_61_axes_0, epsilon = var_1928_to_fp16, x = input_65_cast_fp16)[name = string("normed_61_cast_fp16")];
+            tensor<int32, [2]> var_1941_split_sizes_0 = const()[name = string("op_1941_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_1941_axis_0 = const()[name = string("op_1941_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_1941_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_1941_cast_fp16_1 = split(axis = var_1941_axis_0, split_sizes = var_1941_split_sizes_0, x = normed_61_cast_fp16)[name = string("op_1941_cast_fp16")];
+            tensor<fp16, [256]> const_39_to_fp16 = const()[name = string("const_39_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204346496)))];
+            tensor<fp16, [1, 8, 256]> var_1944_cast_fp16 = mul(x = var_1941_cast_fp16_0, y = const_39_to_fp16)[name = string("op_1944_cast_fp16")];
+            tensor<int32, [4]> var_1950 = const()[name = string("op_1950"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_19 = reshape(shape = var_1950, x = var_1944_cast_fp16)[name = string("q_19")];
+            tensor<fp16, [1, 8, 1, 256]> var_1952 = mul(x = q_19, y = cos_1)[name = string("op_1952")];
+            tensor<int32, [2]> var_1953_split_sizes_0 = const()[name = string("op_1953_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_1953_axis_0 = const()[name = string("op_1953_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_1953_0, tensor<fp16, [1, 8, 1, 128]> var_1953_1 = split(axis = var_1953_axis_0, split_sizes = var_1953_split_sizes_0, x = q_19)[name = string("op_1953")];
+            fp16 const_40_promoted = const()[name = string("const_40_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_1955 = mul(x = var_1953_1, y = const_40_promoted)[name = string("op_1955")];
+            int32 var_1957 = const()[name = string("op_1957"), val = int32(-1)];
+            bool var_1958_interleave_0 = const()[name = string("op_1958_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_1958 = concat(axis = var_1957, interleave = var_1958_interleave_0, values = (var_1955, var_1953_0))[name = string("op_1958")];
+            tensor<fp16, [1, 8, 1, 256]> var_1959 = mul(x = var_1958, y = sin_1)[name = string("op_1959")];
+            tensor<fp16, [1, 8, 1, 256]> q_23 = add(x = var_1952, y = var_1959)[name = string("q_23")];
+            string var_1972_pad_type_0 = const()[name = string("op_1972_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_1972_strides_0 = const()[name = string("op_1972_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_1972_pad_0 = const()[name = string("op_1972_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_1972_dilations_0 = const()[name = string("op_1972_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_1972_groups_0 = const()[name = string("op_1972_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_1972 = conv(dilations = var_1972_dilations_0, groups = var_1972_groups_0, pad = var_1972_pad_0, pad_type = var_1972_pad_type_0, strides = var_1972_strides_0, weight = layers_2_self_attn_k_proj_weight_palettized, x = var_1887_cast_fp16)[name = string("op_1972")];
+            tensor<int32, [4]> var_1977 = const()[name = string("op_1977"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_1978 = reshape(shape = var_1977, x = var_1972)[name = string("op_1978")];
+            tensor<int32, [4]> var_1983 = const()[name = string("op_1983"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_2000_pad_type_0 = const()[name = string("op_2000_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2000_strides_0 = const()[name = string("op_2000_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2000_pad_0 = const()[name = string("op_2000_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2000_dilations_0 = const()[name = string("op_2000_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2000_groups_0 = const()[name = string("op_2000_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2000 = conv(dilations = var_2000_dilations_0, groups = var_2000_groups_0, pad = var_2000_pad_0, pad_type = var_2000_pad_type_0, strides = var_2000_strides_0, weight = layers_2_self_attn_v_proj_weight_palettized, x = var_1887_cast_fp16)[name = string("op_2000")];
+            tensor<int32, [4]> var_2005 = const()[name = string("op_2005"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_2006 = reshape(shape = var_2005, x = var_2000)[name = string("op_2006")];
+            tensor<int32, [4]> var_2011 = const()[name = string("op_2011"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_2021 = const()[name = string("op_2021"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_1984 = transpose(perm = var_1983, x = var_1978)[name = string("transpose_147")];
+            tensor<fp16, [1, 1, 256]> x_67 = reshape(shape = var_2021, x = var_1984)[name = string("x_67")];
+            int32 var_2027 = const()[name = string("op_2027"), val = int32(-1)];
+            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_2033_cast_fp16 = mul(x = x_67, y = const_41_promoted_to_fp16)[name = string("op_2033_cast_fp16")];
+            bool input_67_interleave_0 = const()[name = string("input_67_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_67_cast_fp16 = concat(axis = var_2027, interleave = input_67_interleave_0, values = (x_67, var_2033_cast_fp16))[name = string("input_67_cast_fp16")];
+            tensor<int32, [1]> normed_65_axes_0 = const()[name = string("normed_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2025_to_fp16 = const()[name = string("op_2025_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_65_cast_fp16 = layer_norm(axes = normed_65_axes_0, epsilon = var_2025_to_fp16, x = input_67_cast_fp16)[name = string("normed_65_cast_fp16")];
+            tensor<int32, [2]> var_2038_split_sizes_0 = const()[name = string("op_2038_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_2038_axis_0 = const()[name = string("op_2038_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_2038_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_2038_cast_fp16_1 = split(axis = var_2038_axis_0, split_sizes = var_2038_split_sizes_0, x = normed_65_cast_fp16)[name = string("op_2038_cast_fp16")];
+            tensor<fp16, [256]> const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204347072)))];
+            tensor<fp16, [1, 1, 256]> var_2041_cast_fp16 = mul(x = var_2038_cast_fp16_0, y = const_42_to_fp16)[name = string("op_2041_cast_fp16")];
+            tensor<int32, [4]> var_2047 = const()[name = string("op_2047"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_21 = reshape(shape = var_2047, x = var_2041_cast_fp16)[name = string("q_21")];
+            fp16 var_2054_promoted_to_fp16 = const()[name = string("op_2054_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_2012 = transpose(perm = var_2011, x = var_2006)[name = string("transpose_146")];
+            tensor<fp16, [1, 1, 1, 256]> var_2055_cast_fp16 = pow(x = var_2012, y = var_2054_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
+            tensor<int32, [1]> var_2060_axes_0 = const()[name = string("op_2060_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_2060_keep_dims_0 = const()[name = string("op_2060_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_2060_cast_fp16 = reduce_mean(axes = var_2060_axes_0, keep_dims = var_2060_keep_dims_0, x = var_2055_cast_fp16)[name = string("op_2060_cast_fp16")];
+            fp16 var_2062_to_fp16 = const()[name = string("op_2062_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_5_cast_fp16 = add(x = var_2060_cast_fp16, y = var_2062_to_fp16)[name = string("mean_sq_5_cast_fp16")];
+            fp16 var_2069_to_fp16 = const()[name = string("op_2069_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_2070_cast_fp16 = pow(x = mean_sq_5_cast_fp16, y = var_2069_to_fp16)[name = string("op_2070_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_2071_cast_fp16 = mul(x = var_2012, y = var_2070_cast_fp16)[name = string("op_2071_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_2077 = mul(x = q_21, y = cos_1)[name = string("op_2077")];
+            tensor<int32, [2]> var_2078_split_sizes_0 = const()[name = string("op_2078_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_2078_axis_0 = const()[name = string("op_2078_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_2078_0, tensor<fp16, [1, 1, 1, 128]> var_2078_1 = split(axis = var_2078_axis_0, split_sizes = var_2078_split_sizes_0, x = q_21)[name = string("op_2078")];
+            fp16 const_43_promoted = const()[name = string("const_43_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_2080 = mul(x = var_2078_1, y = const_43_promoted)[name = string("op_2080")];
+            int32 var_2082 = const()[name = string("op_2082"), val = int32(-1)];
+            bool var_2083_interleave_0 = const()[name = string("op_2083_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_2083 = concat(axis = var_2082, interleave = var_2083_interleave_0, values = (var_2080, var_2078_0))[name = string("op_2083")];
+            tensor<fp16, [1, 1, 1, 256]> var_2084 = mul(x = var_2083, y = sin_1)[name = string("op_2084")];
+            tensor<fp16, [1, 1, 1, 256]> input_69 = add(x = var_2077, y = var_2084)[name = string("input_69")];
+            tensor<int32, [4]> var_2089_begin_0 = const()[name = string("op_2089_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> var_2089_end_0 = const()[name = string("op_2089_end_0"), val = tensor<int32, [4]>([3, 1, 512, 512])];
+            tensor<bool, [4]> var_2089_end_mask_0 = const()[name = string("op_2089_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_2089_squeeze_mask_0 = const()[name = string("op_2089_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_2089_cast_fp16 = slice_by_index(begin = var_2089_begin_0, end = var_2089_end_0, end_mask = var_2089_end_mask_0, squeeze_mask = var_2089_squeeze_mask_0, x = coreml_update_state_27)[name = string("op_2089_cast_fp16")];
+            tensor<int32, [1]> K_c_5_axes_0 = const()[name = string("K_c_5_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_5_cast_fp16 = expand_dims(axes = K_c_5_axes_0, x = var_2089_cast_fp16)[name = string("K_c_5_cast_fp16")];
+            tensor<int32, [4]> var_2094_begin_0 = const()[name = string("op_2094_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> var_2094_end_0 = const()[name = string("op_2094_end_0"), val = tensor<int32, [4]>([15, 1, 512, 512])];
+            tensor<bool, [4]> var_2094_end_mask_0 = const()[name = string("op_2094_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_2094_squeeze_mask_0 = const()[name = string("op_2094_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = coreml_update_state_27)[name = string("op_2094_cast_fp16")];
+            tensor<int32, [1]> V_c_5_axes_0 = const()[name = string("V_c_5_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_5_cast_fp16 = expand_dims(axes = V_c_5_axes_0, x = var_2094_cast_fp16)[name = string("V_c_5_cast_fp16")];
+            tensor<int32, [8]> kp_5_pad_0 = const()[name = string("kp_5_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_5_mode_0 = const()[name = string("kp_5_mode_0"), val = string("constant")];
+            fp16 const_44_to_fp16 = const()[name = string("const_44_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_5_cast_fp16 = pad(constant_val = const_44_to_fp16, mode = kp_5_mode_0, pad = kp_5_pad_0, x = input_69)[name = string("kp_5_cast_fp16")];
+            tensor<int32, [8]> vp_5_pad_0 = const()[name = string("vp_5_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_5_mode_0 = const()[name = string("vp_5_mode_0"), val = string("constant")];
+            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_5_cast_fp16 = pad(constant_val = const_45_to_fp16, mode = vp_5_mode_0, pad = vp_5_pad_0, x = var_2071_cast_fp16)[name = string("vp_5_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2112_cast_fp16 = mul(x = K_c_5_cast_fp16, y = var_1005_cast_fp16)[name = string("op_2112_cast_fp16")];
+            tensor<int32, [4]> var_2113_reps_0 = const()[name = string("op_2113_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_2113_cast_fp16 = tile(reps = var_2113_reps_0, x = kp_5_cast_fp16)[name = string("op_2113_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2114_cast_fp16 = mul(x = var_2113_cast_fp16, y = update_mask)[name = string("op_2114_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_5_cast_fp16 = add(x = var_2112_cast_fp16, y = var_2114_cast_fp16)[name = string("K_n_5_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2120_cast_fp16 = mul(x = V_c_5_cast_fp16, y = var_1005_cast_fp16)[name = string("op_2120_cast_fp16")];
+            tensor<int32, [4]> var_2121_reps_0 = const()[name = string("op_2121_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_2121_cast_fp16 = tile(reps = var_2121_reps_0, x = vp_5_cast_fp16)[name = string("op_2121_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2122_cast_fp16 = mul(x = var_2121_cast_fp16, y = update_mask)[name = string("op_2122_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_5_cast_fp16 = add(x = var_2120_cast_fp16, y = var_2122_cast_fp16)[name = string("V_n_5_cast_fp16")];
+            tensor<int32, [1]> var_2126_axes_0 = const()[name = string("op_2126_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_2126_cast_fp16 = squeeze(axes = var_2126_axes_0, x = K_n_5_cast_fp16)[name = string("op_2126_cast_fp16")];
+            tensor<int32, [4]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_5_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_5_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_5_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_16, begin_mask = kv_cache_0_internal_tensor_assign_5_begin_mask_0, end = concat_17, end_mask = kv_cache_0_internal_tensor_assign_5_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_5_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_5_stride_0, update = var_2126_cast_fp16, x = coreml_update_state_27)[name = string("kv_cache_0_internal_tensor_assign_5_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_5_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_28_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_28 = read_state(input = kv_cache_0)[name = string("coreml_update_state_28")];
+            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_2133_cast_fp16 = squeeze(axes = var_2133_axes_0, x = V_n_5_cast_fp16)[name = string("op_2133_cast_fp16")];
+            tensor<int32, [4]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_6_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_6_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_6_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_18, begin_mask = kv_cache_0_internal_tensor_assign_6_begin_mask_0, end = concat_19, end_mask = kv_cache_0_internal_tensor_assign_6_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_6_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_6_stride_0, update = var_2133_cast_fp16, x = coreml_update_state_28)[name = string("kv_cache_0_internal_tensor_assign_6_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_6_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_29_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_29 = read_state(input = kv_cache_0)[name = string("coreml_update_state_29")];
+            tensor<int32, [4]> var_2143_begin_0 = const()[name = string("op_2143_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2143_end_0 = const()[name = string("op_2143_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_2143_end_mask_0 = const()[name = string("op_2143_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_2143_cast_fp16 = slice_by_index(begin = var_2143_begin_0, end = var_2143_end_0, end_mask = var_2143_end_mask_0, x = K_n_5_cast_fp16)[name = string("op_2143_cast_fp16")];
+            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = var_2143_cast_fp16)[name = string("transpose_145")];
+            tensor<fp16, [8, 1, 512, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
+            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_8_cast_fp16 = reshape(shape = concat_20, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
+            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_144")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_9_cast_fp16 = reshape(shape = concat_21, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
+            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = V_n_5_cast_fp16)[name = string("op_2152_cast_fp16")];
+            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = var_2152_cast_fp16)[name = string("transpose_143")];
+            tensor<fp16, [8, 1, 512, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
+            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_10_cast_fp16 = reshape(shape = concat_22, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
+            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_142")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_11_cast_fp16 = reshape(shape = concat_23, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
+            tensor<int32, [4]> Ve_5_perm_0 = const()[name = string("Ve_5_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_2170_transpose_x_0 = const()[name = string("op_2170_transpose_x_0"), val = bool(false)];
+            bool var_2170_transpose_y_0 = const()[name = string("op_2170_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_141")];
+            tensor<fp16, [1, 8, 1, 512]> var_2170_cast_fp16 = matmul(transpose_x = var_2170_transpose_x_0, transpose_y = var_2170_transpose_y_0, x = q_23, y = transpose_50_cast_fp16)[name = string("op_2170_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_2177_cast_fp16 = add(x = var_2170_cast_fp16, y = causal_mask)[name = string("op_2177_cast_fp16")];
+            int32 var_2178 = const()[name = string("op_2178"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_2180_cast_fp16 = softmax(axis = var_2178, x = var_2177_cast_fp16)[name = string("op_2180_cast_fp16")];
+            bool var_2196_transpose_x_0 = const()[name = string("op_2196_transpose_x_0"), val = bool(false)];
+            bool var_2196_transpose_y_0 = const()[name = string("op_2196_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_5_cast_fp16 = transpose(perm = Ve_5_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_140")];
+            tensor<fp16, [1, 8, 1, 256]> var_2196_cast_fp16 = matmul(transpose_x = var_2196_transpose_x_0, transpose_y = var_2196_transpose_y_0, x = var_2180_cast_fp16, y = Ve_5_cast_fp16)[name = string("op_2196_cast_fp16")];
+            tensor<int32, [4]> var_2206 = const()[name = string("op_2206"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2213 = const()[name = string("op_2213"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_2207 = transpose(perm = var_2206, x = var_2196_cast_fp16)[name = string("transpose_139")];
+            tensor<fp16, [1, 1, 2048]> var_2214 = reshape(shape = var_2213, x = var_2207)[name = string("op_2214")];
+            tensor<int32, [3]> var_2218 = const()[name = string("op_2218"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_2_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204347648))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205920576))))[name = string("squeeze_2_palettized")];
+            string var_2234_pad_type_0 = const()[name = string("op_2234_pad_type_0"), val = string("valid")];
+            int32 var_2234_groups_0 = const()[name = string("op_2234_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_2234_strides_0 = const()[name = string("op_2234_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_2234_pad_0 = const()[name = string("op_2234_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_2234_dilations_0 = const()[name = string("op_2234_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_2219 = transpose(perm = var_2218, x = var_2214)[name = string("transpose_138")];
+            tensor<fp16, [1, 1536, 1]> var_2234 = conv(dilations = var_2234_dilations_0, groups = var_2234_groups_0, pad = var_2234_pad_0, pad_type = var_2234_pad_type_0, strides = var_2234_strides_0, weight = squeeze_2_palettized, x = var_2219)[name = string("op_2234")];
+            tensor<int32, [3]> var_2238 = const()[name = string("op_2238"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_2244 = const()[name = string("op_2244"), val = int32(-1)];
+            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_73 = transpose(perm = var_2238, x = var_2234)[name = string("transpose_137")];
+            tensor<fp16, [1, 1, 1536]> var_2250_cast_fp16 = mul(x = x_73, y = const_46_promoted_to_fp16)[name = string("op_2250_cast_fp16")];
+            bool input_75_interleave_0 = const()[name = string("input_75_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_75_cast_fp16 = concat(axis = var_2244, interleave = input_75_interleave_0, values = (x_73, var_2250_cast_fp16))[name = string("input_75_cast_fp16")];
+            tensor<int32, [1]> normed_69_axes_0 = const()[name = string("normed_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2242_to_fp16 = const()[name = string("op_2242_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_2242_to_fp16, x = input_75_cast_fp16)[name = string("normed_69_cast_fp16")];
+            tensor<int32, [2]> var_2255_split_sizes_0 = const()[name = string("op_2255_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2255_axis_0 = const()[name = string("op_2255_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2255_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2255_cast_fp16_1 = split(axis = var_2255_axis_0, split_sizes = var_2255_split_sizes_0, x = normed_69_cast_fp16)[name = string("op_2255_cast_fp16")];
+            tensor<fp16, [1536]> const_47_to_fp16 = const()[name = string("const_47_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205922176)))];
+            tensor<fp16, [1, 1, 1536]> var_2258_cast_fp16 = mul(x = var_2255_cast_fp16_0, y = const_47_to_fp16)[name = string("op_2258_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_77_cast_fp16 = add(x = x_59_cast_fp16, y = var_2258_cast_fp16)[name = string("x_77_cast_fp16")];
+            int32 var_2265 = const()[name = string("op_2265"), val = int32(-1)];
+            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2271_cast_fp16 = mul(x = x_77_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_2271_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_2265, interleave = input_77_interleave_0, values = (x_77_cast_fp16, var_2271_cast_fp16))[name = string("input_77_cast_fp16")];
+            tensor<int32, [1]> normed_73_axes_0 = const()[name = string("normed_73_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2263_to_fp16 = const()[name = string("op_2263_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_73_cast_fp16 = layer_norm(axes = normed_73_axes_0, epsilon = var_2263_to_fp16, x = input_77_cast_fp16)[name = string("normed_73_cast_fp16")];
+            tensor<int32, [2]> var_2276_split_sizes_0 = const()[name = string("op_2276_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2276_axis_0 = const()[name = string("op_2276_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2276_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2276_cast_fp16_1 = split(axis = var_2276_axis_0, split_sizes = var_2276_split_sizes_0, x = normed_73_cast_fp16)[name = string("op_2276_cast_fp16")];
+            tensor<fp16, [1536]> const_49_to_fp16 = const()[name = string("const_49_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205925312)))];
+            tensor<fp16, [1, 1, 1536]> var_2279_cast_fp16 = mul(x = var_2276_cast_fp16_0, y = const_49_to_fp16)[name = string("op_2279_cast_fp16")];
+            tensor<int32, [3]> var_2292 = const()[name = string("op_2292"), 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_2293 = transpose(perm = var_2292, x = var_2279_cast_fp16)[name = string("transpose_136")];
+            tensor<fp16, [1, 1536, 1, 1]> input_79 = expand_dims(axes = input_79_axes_0, x = var_2293)[name = string("input_79")];
+            string var_2306_pad_type_0 = const()[name = string("op_2306_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2306_strides_0 = const()[name = string("op_2306_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2306_pad_0 = const()[name = string("op_2306_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2306_dilations_0 = const()[name = string("op_2306_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2306_groups_0 = const()[name = string("op_2306_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_2306 = conv(dilations = var_2306_dilations_0, groups = var_2306_groups_0, pad = var_2306_pad_0, pad_type = var_2306_pad_type_0, strides = var_2306_strides_0, weight = layers_2_mlp_gate_proj_weight_palettized, x = input_79)[name = string("op_2306")];
+            string var_2308_mode_0 = const()[name = string("op_2308_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_2308 = gelu(mode = var_2308_mode_0, x = var_2306)[name = string("op_2308")];
+            string var_2319_pad_type_0 = const()[name = string("op_2319_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2319_strides_0 = const()[name = string("op_2319_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2319_pad_0 = const()[name = string("op_2319_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2319_dilations_0 = const()[name = string("op_2319_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2319_groups_0 = const()[name = string("op_2319_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_2319 = conv(dilations = var_2319_dilations_0, groups = var_2319_groups_0, pad = var_2319_pad_0, pad_type = var_2319_pad_type_0, strides = var_2319_strides_0, weight = layers_2_mlp_up_proj_weight_palettized, x = input_79)[name = string("op_2319")];
+            tensor<fp16, [1, 6144, 1, 1]> input_81 = mul(x = var_2308, y = var_2319)[name = string("input_81")];
+            string var_2331_pad_type_0 = const()[name = string("op_2331_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2331_strides_0 = const()[name = string("op_2331_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2331_pad_0 = const()[name = string("op_2331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2331_dilations_0 = const()[name = string("op_2331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2331_groups_0 = const()[name = string("op_2331_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_2331 = conv(dilations = var_2331_dilations_0, groups = var_2331_groups_0, pad = var_2331_pad_0, pad_type = var_2331_pad_type_0, strides = var_2331_strides_0, weight = layers_2_mlp_down_proj_weight_palettized, x = input_81)[name = string("op_2331")];
+            tensor<int32, [1]> var_2333_axes_0 = const()[name = string("op_2333_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2333 = squeeze(axes = var_2333_axes_0, x = var_2331)[name = string("op_2333")];
+            tensor<int32, [3]> var_2337 = const()[name = string("op_2337"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_2343 = const()[name = string("op_2343"), val = int32(-1)];
+            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_81 = transpose(perm = var_2337, x = var_2333)[name = string("transpose_135")];
+            tensor<fp16, [1, 1, 1536]> var_2349_cast_fp16 = mul(x = x_81, y = const_50_promoted_to_fp16)[name = string("op_2349_cast_fp16")];
+            bool input_83_interleave_0 = const()[name = string("input_83_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_83_cast_fp16 = concat(axis = var_2343, interleave = input_83_interleave_0, values = (x_81, var_2349_cast_fp16))[name = string("input_83_cast_fp16")];
+            tensor<int32, [1]> normed_77_axes_0 = const()[name = string("normed_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2341_to_fp16 = const()[name = string("op_2341_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_77_cast_fp16 = layer_norm(axes = normed_77_axes_0, epsilon = var_2341_to_fp16, x = input_83_cast_fp16)[name = string("normed_77_cast_fp16")];
+            tensor<int32, [2]> var_2354_split_sizes_0 = const()[name = string("op_2354_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2354_axis_0 = const()[name = string("op_2354_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2354_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2354_cast_fp16_1 = split(axis = var_2354_axis_0, split_sizes = var_2354_split_sizes_0, x = normed_77_cast_fp16)[name = string("op_2354_cast_fp16")];
+            tensor<fp16, [1536]> const_51_to_fp16 = const()[name = string("const_51_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205928448)))];
+            tensor<fp16, [1, 1, 1536]> var_2357_cast_fp16 = mul(x = var_2354_cast_fp16_0, y = const_51_to_fp16)[name = string("op_2357_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_41_cast_fp16 = add(x = x_77_cast_fp16, y = var_2357_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_2368 = linear(bias = linear_0_bias_0, weight = layers_2_per_layer_input_gate_weight_palettized, x = hidden_states_41_cast_fp16)[name = string("linear_4")];
+            string gated_5_mode_0 = const()[name = string("gated_5_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_5 = gelu(mode = gated_5_mode_0, x = var_2368)[name = string("gated_5")];
+            tensor<int32, [3]> var_2385_begin_0 = const()[name = string("op_2385_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
+            tensor<int32, [3]> var_2385_end_0 = const()[name = string("op_2385_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
+            tensor<bool, [3]> var_2385_end_mask_0 = const()[name = string("op_2385_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_2385_cast_fp16 = slice_by_index(begin = var_2385_begin_0, end = var_2385_end_0, end_mask = var_2385_end_mask_0, x = per_layer_combined)[name = string("op_2385_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_87_cast_fp16 = mul(x = gated_5, y = var_2385_cast_fp16)[name = string("input_87_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_2_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205931584))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206128256))))[name = string("layers_2_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_87_cast_fp16)[name = string("linear_5_cast_fp16")];
+            int32 var_2394 = const()[name = string("op_2394"), val = int32(-1)];
+            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2400_cast_fp16 = mul(x = linear_5_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2400_cast_fp16")];
+            bool input_89_interleave_0 = const()[name = string("input_89_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_89_cast_fp16 = concat(axis = var_2394, interleave = input_89_interleave_0, values = (linear_5_cast_fp16, var_2400_cast_fp16))[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> normed_81_axes_0 = const()[name = string("normed_81_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2392_to_fp16 = const()[name = string("op_2392_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_81_cast_fp16 = layer_norm(axes = normed_81_axes_0, epsilon = var_2392_to_fp16, x = input_89_cast_fp16)[name = string("normed_81_cast_fp16")];
+            tensor<int32, [2]> var_2405_split_sizes_0 = const()[name = string("op_2405_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2405_axis_0 = const()[name = string("op_2405_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2405_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2405_cast_fp16_1 = split(axis = var_2405_axis_0, split_sizes = var_2405_split_sizes_0, x = normed_81_cast_fp16)[name = string("op_2405_cast_fp16")];
+            tensor<fp16, [1536]> const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206129856)))];
+            tensor<fp16, [1, 1, 1536]> var_2408_cast_fp16 = mul(x = var_2405_cast_fp16_0, y = const_53_to_fp16)[name = string("op_2408_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_45_cast_fp16 = add(x = hidden_states_41_cast_fp16, y = var_2408_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
+            tensor<fp16, [1]> layers_2_layer_scalar_to_fp16 = const()[name = string("layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.96p-1])];
+            tensor<fp16, [1, 1, 1536]> x_89_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = layers_2_layer_scalar_to_fp16)[name = string("x_89_cast_fp16")];
+            int32 var_2416 = const()[name = string("op_2416"), val = int32(-1)];
+            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2422_cast_fp16 = mul(x = x_89_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_2422_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_2416, interleave = input_91_interleave_0, values = (x_89_cast_fp16, var_2422_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [1]> normed_85_axes_0 = const()[name = string("normed_85_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2414_to_fp16 = const()[name = string("op_2414_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_85_cast_fp16 = layer_norm(axes = normed_85_axes_0, epsilon = var_2414_to_fp16, x = input_91_cast_fp16)[name = string("normed_85_cast_fp16")];
+            tensor<int32, [2]> var_2427_split_sizes_0 = const()[name = string("op_2427_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2427_axis_0 = const()[name = string("op_2427_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2427_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2427_cast_fp16_1 = split(axis = var_2427_axis_0, split_sizes = var_2427_split_sizes_0, x = normed_85_cast_fp16)[name = string("op_2427_cast_fp16")];
+            tensor<fp16, [1536]> const_55_to_fp16 = const()[name = string("const_55_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206132992)))];
+            tensor<fp16, [1, 1, 1536]> var_2430_cast_fp16 = mul(x = var_2427_cast_fp16_0, y = const_55_to_fp16)[name = string("op_2430_cast_fp16")];
+            tensor<int32, [3]> var_2438 = const()[name = string("op_2438"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_2441_axes_0 = const()[name = string("op_2441_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2439_cast_fp16 = transpose(perm = var_2438, x = var_2430_cast_fp16)[name = string("transpose_134")];
+            tensor<fp16, [1, 1536, 1, 1]> var_2441_cast_fp16 = expand_dims(axes = var_2441_axes_0, x = var_2439_cast_fp16)[name = string("op_2441_cast_fp16")];
+            string var_2457_pad_type_0 = const()[name = string("op_2457_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2457_strides_0 = const()[name = string("op_2457_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2457_pad_0 = const()[name = string("op_2457_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2457_dilations_0 = const()[name = string("op_2457_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2457_groups_0 = const()[name = string("op_2457_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_2457 = conv(dilations = var_2457_dilations_0, groups = var_2457_groups_0, pad = var_2457_pad_0, pad_type = var_2457_pad_type_0, strides = var_2457_strides_0, weight = layers_3_self_attn_q_proj_weight_palettized, x = var_2441_cast_fp16)[name = string("op_2457")];
+            tensor<int32, [4]> var_2462 = const()[name = string("op_2462"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_2463 = reshape(shape = var_2462, x = var_2457)[name = string("op_2463")];
+            tensor<int32, [4]> var_2468 = const()[name = string("op_2468"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_2478 = const()[name = string("op_2478"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_2469 = transpose(perm = var_2468, x = var_2463)[name = string("transpose_133")];
+            tensor<fp16, [1, 8, 256]> x_93 = reshape(shape = var_2478, x = var_2469)[name = string("x_93")];
+            int32 var_2484 = const()[name = string("op_2484"), val = int32(-1)];
+            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_2490_cast_fp16 = mul(x = x_93, y = const_56_promoted_to_fp16)[name = string("op_2490_cast_fp16")];
+            bool input_95_interleave_0 = const()[name = string("input_95_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_95_cast_fp16 = concat(axis = var_2484, interleave = input_95_interleave_0, values = (x_93, var_2490_cast_fp16))[name = string("input_95_cast_fp16")];
+            tensor<int32, [1]> normed_89_axes_0 = const()[name = string("normed_89_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_89_cast_fp16 = layer_norm(axes = normed_89_axes_0, epsilon = var_2482_to_fp16, x = input_95_cast_fp16)[name = string("normed_89_cast_fp16")];
+            tensor<int32, [2]> var_2495_split_sizes_0 = const()[name = string("op_2495_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_2495_axis_0 = const()[name = string("op_2495_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_2495_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_2495_cast_fp16_1 = split(axis = var_2495_axis_0, split_sizes = var_2495_split_sizes_0, x = normed_89_cast_fp16)[name = string("op_2495_cast_fp16")];
+            tensor<fp16, [256]> const_57_to_fp16 = const()[name = string("const_57_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206136128)))];
+            tensor<fp16, [1, 8, 256]> var_2498_cast_fp16 = mul(x = var_2495_cast_fp16_0, y = const_57_to_fp16)[name = string("op_2498_cast_fp16")];
+            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_27 = reshape(shape = var_2504, x = var_2498_cast_fp16)[name = string("q_27")];
+            tensor<fp16, [1, 8, 1, 256]> var_2506 = mul(x = q_27, y = cos_1)[name = string("op_2506")];
+            tensor<int32, [2]> var_2507_split_sizes_0 = const()[name = string("op_2507_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_2507_axis_0 = const()[name = string("op_2507_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_2507_0, tensor<fp16, [1, 8, 1, 128]> var_2507_1 = split(axis = var_2507_axis_0, split_sizes = var_2507_split_sizes_0, x = q_27)[name = string("op_2507")];
+            fp16 const_58_promoted = const()[name = string("const_58_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_2509 = mul(x = var_2507_1, y = const_58_promoted)[name = string("op_2509")];
+            int32 var_2511 = const()[name = string("op_2511"), val = int32(-1)];
+            bool var_2512_interleave_0 = const()[name = string("op_2512_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_2512 = concat(axis = var_2511, interleave = var_2512_interleave_0, values = (var_2509, var_2507_0))[name = string("op_2512")];
+            tensor<fp16, [1, 8, 1, 256]> var_2513 = mul(x = var_2512, y = sin_1)[name = string("op_2513")];
+            tensor<fp16, [1, 8, 1, 256]> q_31 = add(x = var_2506, y = var_2513)[name = string("q_31")];
+            string var_2526_pad_type_0 = const()[name = string("op_2526_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2526_strides_0 = const()[name = string("op_2526_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2526_pad_0 = const()[name = string("op_2526_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2526_dilations_0 = const()[name = string("op_2526_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2526_groups_0 = const()[name = string("op_2526_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2526 = conv(dilations = var_2526_dilations_0, groups = var_2526_groups_0, pad = var_2526_pad_0, pad_type = var_2526_pad_type_0, strides = var_2526_strides_0, weight = layers_3_self_attn_k_proj_weight_palettized, x = var_2441_cast_fp16)[name = string("op_2526")];
+            tensor<int32, [4]> var_2531 = const()[name = string("op_2531"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_2532 = reshape(shape = var_2531, x = var_2526)[name = string("op_2532")];
+            tensor<int32, [4]> var_2537 = const()[name = string("op_2537"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_2554_pad_type_0 = const()[name = string("op_2554_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2554_strides_0 = const()[name = string("op_2554_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2554_pad_0 = const()[name = string("op_2554_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2554_dilations_0 = const()[name = string("op_2554_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2554_groups_0 = const()[name = string("op_2554_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_2554 = conv(dilations = var_2554_dilations_0, groups = var_2554_groups_0, pad = var_2554_pad_0, pad_type = var_2554_pad_type_0, strides = var_2554_strides_0, weight = layers_3_self_attn_v_proj_weight_palettized, x = var_2441_cast_fp16)[name = string("op_2554")];
+            tensor<int32, [4]> var_2559 = const()[name = string("op_2559"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_2560 = reshape(shape = var_2559, x = var_2554)[name = string("op_2560")];
+            tensor<int32, [4]> var_2565 = const()[name = string("op_2565"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_2575 = const()[name = string("op_2575"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_2538 = transpose(perm = var_2537, x = var_2532)[name = string("transpose_132")];
+            tensor<fp16, [1, 1, 256]> x_97 = reshape(shape = var_2575, x = var_2538)[name = string("x_97")];
+            int32 var_2581 = const()[name = string("op_2581"), val = int32(-1)];
+            fp16 const_59_promoted_to_fp16 = const()[name = string("const_59_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_2587_cast_fp16 = mul(x = x_97, y = const_59_promoted_to_fp16)[name = string("op_2587_cast_fp16")];
+            bool input_97_interleave_0 = const()[name = string("input_97_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_97_cast_fp16 = concat(axis = var_2581, interleave = input_97_interleave_0, values = (x_97, var_2587_cast_fp16))[name = string("input_97_cast_fp16")];
+            tensor<int32, [1]> normed_93_axes_0 = const()[name = string("normed_93_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2579_to_fp16 = const()[name = string("op_2579_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_93_cast_fp16 = layer_norm(axes = normed_93_axes_0, epsilon = var_2579_to_fp16, x = input_97_cast_fp16)[name = string("normed_93_cast_fp16")];
+            tensor<int32, [2]> var_2592_split_sizes_0 = const()[name = string("op_2592_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_2592_axis_0 = const()[name = string("op_2592_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_2592_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_2592_cast_fp16_1 = split(axis = var_2592_axis_0, split_sizes = var_2592_split_sizes_0, x = normed_93_cast_fp16)[name = string("op_2592_cast_fp16")];
+            tensor<fp16, [256]> const_60_to_fp16 = const()[name = string("const_60_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206136704)))];
+            tensor<fp16, [1, 1, 256]> var_2595_cast_fp16 = mul(x = var_2592_cast_fp16_0, y = const_60_to_fp16)[name = string("op_2595_cast_fp16")];
+            tensor<int32, [4]> var_2601 = const()[name = string("op_2601"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_29 = reshape(shape = var_2601, x = var_2595_cast_fp16)[name = string("q_29")];
+            fp16 var_2608_promoted_to_fp16 = const()[name = string("op_2608_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_2566 = transpose(perm = var_2565, x = var_2560)[name = string("transpose_131")];
+            tensor<fp16, [1, 1, 1, 256]> var_2609_cast_fp16 = pow(x = var_2566, y = var_2608_promoted_to_fp16)[name = string("op_2609_cast_fp16")];
+            tensor<int32, [1]> var_2614_axes_0 = const()[name = string("op_2614_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_2614_keep_dims_0 = const()[name = string("op_2614_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_2614_cast_fp16 = reduce_mean(axes = var_2614_axes_0, keep_dims = var_2614_keep_dims_0, x = var_2609_cast_fp16)[name = string("op_2614_cast_fp16")];
+            fp16 var_2616_to_fp16 = const()[name = string("op_2616_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_7_cast_fp16 = add(x = var_2614_cast_fp16, y = var_2616_to_fp16)[name = string("mean_sq_7_cast_fp16")];
+            fp16 var_2623_to_fp16 = const()[name = string("op_2623_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_2624_cast_fp16 = pow(x = mean_sq_7_cast_fp16, y = var_2623_to_fp16)[name = string("op_2624_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_2625_cast_fp16 = mul(x = var_2566, y = var_2624_cast_fp16)[name = string("op_2625_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_2631 = mul(x = q_29, y = cos_1)[name = string("op_2631")];
+            tensor<int32, [2]> var_2632_split_sizes_0 = const()[name = string("op_2632_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_2632_axis_0 = const()[name = string("op_2632_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_2632_0, tensor<fp16, [1, 1, 1, 128]> var_2632_1 = split(axis = var_2632_axis_0, split_sizes = var_2632_split_sizes_0, x = q_29)[name = string("op_2632")];
+            fp16 const_61_promoted = const()[name = string("const_61_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_2634 = mul(x = var_2632_1, y = const_61_promoted)[name = string("op_2634")];
+            int32 var_2636 = const()[name = string("op_2636"), val = int32(-1)];
+            bool var_2637_interleave_0 = const()[name = string("op_2637_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_2637 = concat(axis = var_2636, interleave = var_2637_interleave_0, values = (var_2634, var_2632_0))[name = string("op_2637")];
+            tensor<fp16, [1, 1, 1, 256]> var_2638 = mul(x = var_2637, y = sin_1)[name = string("op_2638")];
+            tensor<fp16, [1, 1, 1, 256]> input_99 = add(x = var_2631, y = var_2638)[name = string("input_99")];
+            tensor<int32, [4]> var_2643_begin_0 = const()[name = string("op_2643_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> var_2643_end_0 = const()[name = string("op_2643_end_0"), val = tensor<int32, [4]>([4, 1, 512, 512])];
+            tensor<bool, [4]> var_2643_end_mask_0 = const()[name = string("op_2643_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_2643_squeeze_mask_0 = const()[name = string("op_2643_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_2643_cast_fp16 = slice_by_index(begin = var_2643_begin_0, end = var_2643_end_0, end_mask = var_2643_end_mask_0, squeeze_mask = var_2643_squeeze_mask_0, x = coreml_update_state_29)[name = string("op_2643_cast_fp16")];
+            tensor<int32, [1]> K_c_7_axes_0 = const()[name = string("K_c_7_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_7_cast_fp16 = expand_dims(axes = K_c_7_axes_0, x = var_2643_cast_fp16)[name = string("K_c_7_cast_fp16")];
+            tensor<int32, [4]> var_2648_begin_0 = const()[name = string("op_2648_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> var_2648_end_0 = const()[name = string("op_2648_end_0"), val = tensor<int32, [4]>([16, 1, 512, 512])];
+            tensor<bool, [4]> var_2648_end_mask_0 = const()[name = string("op_2648_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_2648_squeeze_mask_0 = const()[name = string("op_2648_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_2648_cast_fp16 = slice_by_index(begin = var_2648_begin_0, end = var_2648_end_0, end_mask = var_2648_end_mask_0, squeeze_mask = var_2648_squeeze_mask_0, x = coreml_update_state_29)[name = string("op_2648_cast_fp16")];
+            tensor<int32, [1]> V_c_7_axes_0 = const()[name = string("V_c_7_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_7_cast_fp16 = expand_dims(axes = V_c_7_axes_0, x = var_2648_cast_fp16)[name = string("V_c_7_cast_fp16")];
+            tensor<int32, [8]> kp_7_pad_0 = const()[name = string("kp_7_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_7_mode_0 = const()[name = string("kp_7_mode_0"), val = string("constant")];
+            fp16 const_62_to_fp16 = const()[name = string("const_62_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_7_cast_fp16 = pad(constant_val = const_62_to_fp16, mode = kp_7_mode_0, pad = kp_7_pad_0, x = input_99)[name = string("kp_7_cast_fp16")];
+            tensor<int32, [8]> vp_7_pad_0 = const()[name = string("vp_7_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_7_mode_0 = const()[name = string("vp_7_mode_0"), val = string("constant")];
+            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_7_cast_fp16 = pad(constant_val = const_63_to_fp16, mode = vp_7_mode_0, pad = vp_7_pad_0, x = var_2625_cast_fp16)[name = string("vp_7_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2666_cast_fp16 = mul(x = K_c_7_cast_fp16, y = var_1005_cast_fp16)[name = string("op_2666_cast_fp16")];
+            tensor<int32, [4]> var_2667_reps_0 = const()[name = string("op_2667_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_2667_cast_fp16 = tile(reps = var_2667_reps_0, x = kp_7_cast_fp16)[name = string("op_2667_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2668_cast_fp16 = mul(x = var_2667_cast_fp16, y = update_mask)[name = string("op_2668_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_7_cast_fp16 = add(x = var_2666_cast_fp16, y = var_2668_cast_fp16)[name = string("K_n_7_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2674_cast_fp16 = mul(x = V_c_7_cast_fp16, y = var_1005_cast_fp16)[name = string("op_2674_cast_fp16")];
+            tensor<int32, [4]> var_2675_reps_0 = const()[name = string("op_2675_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_2675_cast_fp16 = tile(reps = var_2675_reps_0, x = vp_7_cast_fp16)[name = string("op_2675_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_2676_cast_fp16 = mul(x = var_2675_cast_fp16, y = update_mask)[name = string("op_2676_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_7_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2676_cast_fp16)[name = string("V_n_7_cast_fp16")];
+            tensor<int32, [1]> var_2680_axes_0 = const()[name = string("op_2680_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_2680_cast_fp16 = squeeze(axes = var_2680_axes_0, x = K_n_7_cast_fp16)[name = string("op_2680_cast_fp16")];
+            tensor<int32, [4]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_7_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_7_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_7_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_24, begin_mask = kv_cache_0_internal_tensor_assign_7_begin_mask_0, end = concat_25, end_mask = kv_cache_0_internal_tensor_assign_7_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_7_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_7_stride_0, update = var_2680_cast_fp16, x = coreml_update_state_29)[name = string("kv_cache_0_internal_tensor_assign_7_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_7_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_30_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_30 = read_state(input = kv_cache_0)[name = string("coreml_update_state_30")];
+            tensor<int32, [1]> var_2687_axes_0 = const()[name = string("op_2687_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_2687_cast_fp16 = squeeze(axes = var_2687_axes_0, x = V_n_7_cast_fp16)[name = string("op_2687_cast_fp16")];
+            tensor<int32, [4]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_8_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_8_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_8_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_26, begin_mask = kv_cache_0_internal_tensor_assign_8_begin_mask_0, end = concat_27, end_mask = kv_cache_0_internal_tensor_assign_8_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_8_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_8_stride_0, update = var_2687_cast_fp16, x = coreml_update_state_30)[name = string("kv_cache_0_internal_tensor_assign_8_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_8_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_31_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_31 = read_state(input = kv_cache_0)[name = string("coreml_update_state_31")];
+            tensor<int32, [4]> var_2697_begin_0 = const()[name = string("op_2697_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2697_end_0 = const()[name = string("op_2697_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_2697_end_mask_0 = const()[name = string("op_2697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_2697_cast_fp16 = slice_by_index(begin = var_2697_begin_0, end = var_2697_end_0, end_mask = var_2697_end_mask_0, x = K_n_7_cast_fp16)[name = string("op_2697_cast_fp16")];
+            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = var_2697_cast_fp16)[name = string("transpose_130")];
+            tensor<fp16, [8, 1, 512, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
+            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_12_cast_fp16 = reshape(shape = concat_28, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
+            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_129")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_13_cast_fp16 = reshape(shape = concat_29, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
+            tensor<int32, [4]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_2706_begin_0 = const()[name = string("op_2706_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2706_end_0 = const()[name = string("op_2706_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_2706_end_mask_0 = const()[name = string("op_2706_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_2706_cast_fp16 = slice_by_index(begin = var_2706_begin_0, end = var_2706_end_0, end_mask = var_2706_end_mask_0, x = V_n_7_cast_fp16)[name = string("op_2706_cast_fp16")];
+            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = var_2706_cast_fp16)[name = string("transpose_128")];
+            tensor<fp16, [8, 1, 512, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
+            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_14_cast_fp16 = reshape(shape = concat_30, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
+            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_127")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_15_cast_fp16 = reshape(shape = concat_31, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
+            tensor<int32, [4]> Ve_7_perm_0 = const()[name = string("Ve_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_2724_transpose_x_0 = const()[name = string("op_2724_transpose_x_0"), val = bool(false)];
+            bool var_2724_transpose_y_0 = const()[name = string("op_2724_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_126")];
+            tensor<fp16, [1, 8, 1, 512]> var_2724_cast_fp16 = matmul(transpose_x = var_2724_transpose_x_0, transpose_y = var_2724_transpose_y_0, x = q_31, y = transpose_51_cast_fp16)[name = string("op_2724_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_2731_cast_fp16 = add(x = var_2724_cast_fp16, y = causal_mask)[name = string("op_2731_cast_fp16")];
+            int32 var_2732 = const()[name = string("op_2732"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_2734_cast_fp16 = softmax(axis = var_2732, x = var_2731_cast_fp16)[name = string("op_2734_cast_fp16")];
+            bool var_2750_transpose_x_0 = const()[name = string("op_2750_transpose_x_0"), val = bool(false)];
+            bool var_2750_transpose_y_0 = const()[name = string("op_2750_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_7_cast_fp16 = transpose(perm = Ve_7_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_125")];
+            tensor<fp16, [1, 8, 1, 256]> var_2750_cast_fp16 = matmul(transpose_x = var_2750_transpose_x_0, transpose_y = var_2750_transpose_y_0, x = var_2734_cast_fp16, y = Ve_7_cast_fp16)[name = string("op_2750_cast_fp16")];
+            tensor<int32, [4]> var_2760 = const()[name = string("op_2760"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2767 = const()[name = string("op_2767"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_2761 = transpose(perm = var_2760, x = var_2750_cast_fp16)[name = string("transpose_124")];
+            tensor<fp16, [1, 1, 2048]> var_2768 = reshape(shape = var_2767, x = var_2761)[name = string("op_2768")];
+            tensor<int32, [3]> var_2772 = const()[name = string("op_2772"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_3_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206137280))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207710208))))[name = string("squeeze_3_palettized")];
+            string var_2788_pad_type_0 = const()[name = string("op_2788_pad_type_0"), val = string("valid")];
+            int32 var_2788_groups_0 = const()[name = string("op_2788_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_2788_strides_0 = const()[name = string("op_2788_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_2788_pad_0 = const()[name = string("op_2788_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_2788_dilations_0 = const()[name = string("op_2788_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_2773 = transpose(perm = var_2772, x = var_2768)[name = string("transpose_123")];
+            tensor<fp16, [1, 1536, 1]> var_2788 = conv(dilations = var_2788_dilations_0, groups = var_2788_groups_0, pad = var_2788_pad_0, pad_type = var_2788_pad_type_0, strides = var_2788_strides_0, weight = squeeze_3_palettized, x = var_2773)[name = string("op_2788")];
+            tensor<int32, [3]> var_2792 = const()[name = string("op_2792"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_2798 = const()[name = string("op_2798"), val = int32(-1)];
+            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_103 = transpose(perm = var_2792, x = var_2788)[name = string("transpose_122")];
+            tensor<fp16, [1, 1, 1536]> var_2804_cast_fp16 = mul(x = x_103, y = const_64_promoted_to_fp16)[name = string("op_2804_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_2798, interleave = input_105_interleave_0, values = (x_103, var_2804_cast_fp16))[name = string("input_105_cast_fp16")];
+            tensor<int32, [1]> normed_97_axes_0 = const()[name = string("normed_97_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2796_to_fp16 = const()[name = string("op_2796_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_97_cast_fp16 = layer_norm(axes = normed_97_axes_0, epsilon = var_2796_to_fp16, x = input_105_cast_fp16)[name = string("normed_97_cast_fp16")];
+            tensor<int32, [2]> var_2809_split_sizes_0 = const()[name = string("op_2809_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2809_axis_0 = const()[name = string("op_2809_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2809_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2809_cast_fp16_1 = split(axis = var_2809_axis_0, split_sizes = var_2809_split_sizes_0, x = normed_97_cast_fp16)[name = string("op_2809_cast_fp16")];
+            tensor<fp16, [1536]> const_65_to_fp16 = const()[name = string("const_65_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207711808)))];
+            tensor<fp16, [1, 1, 1536]> var_2812_cast_fp16 = mul(x = var_2809_cast_fp16_0, y = const_65_to_fp16)[name = string("op_2812_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_107_cast_fp16 = add(x = x_89_cast_fp16, y = var_2812_cast_fp16)[name = string("x_107_cast_fp16")];
+            int32 var_2819 = const()[name = string("op_2819"), 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_2825_cast_fp16 = mul(x = x_107_cast_fp16, y = const_66_promoted_to_fp16)[name = string("op_2825_cast_fp16")];
+            bool input_107_interleave_0 = const()[name = string("input_107_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_107_cast_fp16 = concat(axis = var_2819, interleave = input_107_interleave_0, values = (x_107_cast_fp16, var_2825_cast_fp16))[name = string("input_107_cast_fp16")];
+            tensor<int32, [1]> normed_101_axes_0 = const()[name = string("normed_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2817_to_fp16 = const()[name = string("op_2817_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_101_cast_fp16 = layer_norm(axes = normed_101_axes_0, epsilon = var_2817_to_fp16, x = input_107_cast_fp16)[name = string("normed_101_cast_fp16")];
+            tensor<int32, [2]> var_2830_split_sizes_0 = const()[name = string("op_2830_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2830_axis_0 = const()[name = string("op_2830_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2830_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2830_cast_fp16_1 = split(axis = var_2830_axis_0, split_sizes = var_2830_split_sizes_0, x = normed_101_cast_fp16)[name = string("op_2830_cast_fp16")];
+            tensor<fp16, [1536]> const_67_to_fp16 = const()[name = string("const_67_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207714944)))];
+            tensor<fp16, [1, 1, 1536]> var_2833_cast_fp16 = mul(x = var_2830_cast_fp16_0, y = const_67_to_fp16)[name = string("op_2833_cast_fp16")];
+            tensor<int32, [3]> var_2846 = const()[name = string("op_2846"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_109_axes_0 = const()[name = string("input_109_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2847 = transpose(perm = var_2846, x = var_2833_cast_fp16)[name = string("transpose_121")];
+            tensor<fp16, [1, 1536, 1, 1]> input_109 = expand_dims(axes = input_109_axes_0, x = var_2847)[name = string("input_109")];
+            string var_2860_pad_type_0 = const()[name = string("op_2860_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2860_strides_0 = const()[name = string("op_2860_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2860_pad_0 = const()[name = string("op_2860_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2860_dilations_0 = const()[name = string("op_2860_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2860_groups_0 = const()[name = string("op_2860_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_2860 = conv(dilations = var_2860_dilations_0, groups = var_2860_groups_0, pad = var_2860_pad_0, pad_type = var_2860_pad_type_0, strides = var_2860_strides_0, weight = layers_3_mlp_gate_proj_weight_palettized, x = input_109)[name = string("op_2860")];
+            string var_2862_mode_0 = const()[name = string("op_2862_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_2862 = gelu(mode = var_2862_mode_0, x = var_2860)[name = string("op_2862")];
+            string var_2873_pad_type_0 = const()[name = string("op_2873_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2873_strides_0 = const()[name = string("op_2873_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2873_pad_0 = const()[name = string("op_2873_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2873_dilations_0 = const()[name = string("op_2873_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2873_groups_0 = const()[name = string("op_2873_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> 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 = layers_3_mlp_up_proj_weight_palettized, x = input_109)[name = string("op_2873")];
+            tensor<fp16, [1, 6144, 1, 1]> input_111 = mul(x = var_2862, y = var_2873)[name = string("input_111")];
+            string var_2885_pad_type_0 = const()[name = string("op_2885_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_2885_strides_0 = const()[name = string("op_2885_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_2885_pad_0 = const()[name = string("op_2885_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_2885_dilations_0 = const()[name = string("op_2885_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_2885_groups_0 = const()[name = string("op_2885_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_2885 = conv(dilations = var_2885_dilations_0, groups = var_2885_groups_0, pad = var_2885_pad_0, pad_type = var_2885_pad_type_0, strides = var_2885_strides_0, weight = layers_3_mlp_down_proj_weight_palettized, x = input_111)[name = string("op_2885")];
+            tensor<int32, [1]> var_2887_axes_0 = const()[name = string("op_2887_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2887 = squeeze(axes = var_2887_axes_0, x = var_2885)[name = string("op_2887")];
+            tensor<int32, [3]> var_2891 = const()[name = string("op_2891"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_2897 = const()[name = string("op_2897"), val = int32(-1)];
+            fp16 const_68_promoted_to_fp16 = const()[name = string("const_68_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_111 = transpose(perm = var_2891, x = var_2887)[name = string("transpose_120")];
+            tensor<fp16, [1, 1, 1536]> var_2903_cast_fp16 = mul(x = x_111, y = const_68_promoted_to_fp16)[name = string("op_2903_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_2897, interleave = input_113_interleave_0, values = (x_111, var_2903_cast_fp16))[name = string("input_113_cast_fp16")];
+            tensor<int32, [1]> normed_105_axes_0 = const()[name = string("normed_105_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2895_to_fp16 = const()[name = string("op_2895_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_105_cast_fp16 = layer_norm(axes = normed_105_axes_0, epsilon = var_2895_to_fp16, x = input_113_cast_fp16)[name = string("normed_105_cast_fp16")];
+            tensor<int32, [2]> var_2908_split_sizes_0 = const()[name = string("op_2908_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2908_axis_0 = const()[name = string("op_2908_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2908_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2908_cast_fp16_1 = split(axis = var_2908_axis_0, split_sizes = var_2908_split_sizes_0, x = normed_105_cast_fp16)[name = string("op_2908_cast_fp16")];
+            tensor<fp16, [1536]> const_69_to_fp16 = const()[name = string("const_69_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207718080)))];
+            tensor<fp16, [1, 1, 1536]> var_2911_cast_fp16 = mul(x = var_2908_cast_fp16_0, y = const_69_to_fp16)[name = string("op_2911_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_55_cast_fp16 = add(x = x_107_cast_fp16, y = var_2911_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_2922 = linear(bias = linear_0_bias_0, weight = layers_3_per_layer_input_gate_weight_palettized, x = hidden_states_55_cast_fp16)[name = string("linear_6")];
+            string gated_7_mode_0 = const()[name = string("gated_7_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_7 = gelu(mode = gated_7_mode_0, x = var_2922)[name = string("gated_7")];
+            tensor<int32, [3]> var_2939_begin_0 = const()[name = string("op_2939_begin_0"), val = tensor<int32, [3]>([0, 0, 768])];
+            tensor<int32, [3]> var_2939_end_0 = const()[name = string("op_2939_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
+            tensor<bool, [3]> var_2939_end_mask_0 = const()[name = string("op_2939_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_2939_cast_fp16 = slice_by_index(begin = var_2939_begin_0, end = var_2939_end_0, end_mask = var_2939_end_mask_0, x = per_layer_combined)[name = string("op_2939_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_117_cast_fp16 = mul(x = gated_7, y = var_2939_cast_fp16)[name = string("input_117_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_3_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207721216))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207917888))))[name = string("layers_3_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_117_cast_fp16)[name = string("linear_7_cast_fp16")];
+            int32 var_2948 = const()[name = string("op_2948"), val = int32(-1)];
+            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2954_cast_fp16 = mul(x = linear_7_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2954_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_2948, interleave = input_119_interleave_0, values = (linear_7_cast_fp16, var_2954_cast_fp16))[name = string("input_119_cast_fp16")];
+            tensor<int32, [1]> normed_109_axes_0 = const()[name = string("normed_109_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2946_to_fp16 = const()[name = string("op_2946_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_109_cast_fp16 = layer_norm(axes = normed_109_axes_0, epsilon = var_2946_to_fp16, x = input_119_cast_fp16)[name = string("normed_109_cast_fp16")];
+            tensor<int32, [2]> var_2959_split_sizes_0 = const()[name = string("op_2959_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2959_axis_0 = const()[name = string("op_2959_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2959_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2959_cast_fp16_1 = split(axis = var_2959_axis_0, split_sizes = var_2959_split_sizes_0, x = normed_109_cast_fp16)[name = string("op_2959_cast_fp16")];
+            tensor<fp16, [1536]> const_71_to_fp16 = const()[name = string("const_71_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207919488)))];
+            tensor<fp16, [1, 1, 1536]> var_2962_cast_fp16 = mul(x = var_2959_cast_fp16_0, y = const_71_to_fp16)[name = string("op_2962_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = var_2962_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
+            tensor<fp16, [1]> layers_3_layer_scalar_to_fp16 = const()[name = string("layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.26p-2])];
+            tensor<fp16, [1, 1, 1536]> x_119_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = layers_3_layer_scalar_to_fp16)[name = string("x_119_cast_fp16")];
+            int32 var_2970 = const()[name = string("op_2970"), val = int32(-1)];
+            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_2976_cast_fp16 = mul(x = x_119_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2976_cast_fp16")];
+            bool input_121_interleave_0 = const()[name = string("input_121_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_121_cast_fp16 = concat(axis = var_2970, interleave = input_121_interleave_0, values = (x_119_cast_fp16, var_2976_cast_fp16))[name = string("input_121_cast_fp16")];
+            tensor<int32, [1]> normed_113_axes_0 = const()[name = string("normed_113_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_2968_to_fp16 = const()[name = string("op_2968_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_113_cast_fp16 = layer_norm(axes = normed_113_axes_0, epsilon = var_2968_to_fp16, x = input_121_cast_fp16)[name = string("normed_113_cast_fp16")];
+            tensor<int32, [2]> var_2981_split_sizes_0 = const()[name = string("op_2981_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_2981_axis_0 = const()[name = string("op_2981_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_2981_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_2981_cast_fp16_1 = split(axis = var_2981_axis_0, split_sizes = var_2981_split_sizes_0, x = normed_113_cast_fp16)[name = string("op_2981_cast_fp16")];
+            tensor<fp16, [1536]> const_73_to_fp16 = const()[name = string("const_73_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207922624)))];
+            tensor<fp16, [1, 1, 1536]> var_2984_cast_fp16 = mul(x = var_2981_cast_fp16_0, y = const_73_to_fp16)[name = string("op_2984_cast_fp16")];
+            tensor<int32, [3]> var_2992 = const()[name = string("op_2992"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_2995_axes_0 = const()[name = string("op_2995_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_2993_cast_fp16 = transpose(perm = var_2992, x = var_2984_cast_fp16)[name = string("transpose_119")];
+            tensor<fp16, [1, 1536, 1, 1]> var_2995_cast_fp16 = expand_dims(axes = var_2995_axes_0, x = var_2993_cast_fp16)[name = string("op_2995_cast_fp16")];
+            string var_3011_pad_type_0 = const()[name = string("op_3011_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3011_strides_0 = const()[name = string("op_3011_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3011_pad_0 = const()[name = string("op_3011_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3011_dilations_0 = const()[name = string("op_3011_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3011_groups_0 = const()[name = string("op_3011_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 4096, 1, 1]> var_3011 = conv(dilations = var_3011_dilations_0, groups = var_3011_groups_0, pad = var_3011_pad_0, pad_type = var_3011_pad_type_0, strides = var_3011_strides_0, weight = layers_4_self_attn_q_proj_weight_palettized, x = var_2995_cast_fp16)[name = string("op_3011")];
+            tensor<int32, [4]> var_3016 = const()[name = string("op_3016"), val = tensor<int32, [4]>([1, 8, 512, 1])];
+            tensor<fp16, [1, 8, 512, 1]> var_3017 = reshape(shape = var_3016, x = var_3011)[name = string("op_3017")];
+            tensor<int32, [4]> var_3022 = const()[name = string("op_3022"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_3032 = const()[name = string("op_3032"), val = tensor<int32, [3]>([1, 8, 512])];
+            tensor<fp16, [1, 8, 1, 512]> var_3023 = transpose(perm = var_3022, x = var_3017)[name = string("transpose_118")];
+            tensor<fp16, [1, 8, 512]> x_123 = reshape(shape = var_3032, x = var_3023)[name = string("x_123")];
+            int32 var_3038 = const()[name = string("op_3038"), val = int32(-1)];
+            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 512]> var_3044_cast_fp16 = mul(x = x_123, y = const_74_promoted_to_fp16)[name = string("op_3044_cast_fp16")];
+            bool input_125_interleave_0 = const()[name = string("input_125_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1024]> input_125_cast_fp16 = concat(axis = var_3038, interleave = input_125_interleave_0, values = (x_123, var_3044_cast_fp16))[name = string("input_125_cast_fp16")];
+            tensor<int32, [1]> normed_117_axes_0 = const()[name = string("normed_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 1024]> normed_117_cast_fp16 = layer_norm(axes = normed_117_axes_0, epsilon = var_3036_to_fp16, x = input_125_cast_fp16)[name = string("normed_117_cast_fp16")];
+            tensor<int32, [2]> var_3049_split_sizes_0 = const()[name = string("op_3049_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            int32 var_3049_axis_0 = const()[name = string("op_3049_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 512]> var_3049_cast_fp16_0, tensor<fp16, [1, 8, 512]> var_3049_cast_fp16_1 = split(axis = var_3049_axis_0, split_sizes = var_3049_split_sizes_0, x = normed_117_cast_fp16)[name = string("op_3049_cast_fp16")];
+            tensor<fp16, [512]> const_75_to_fp16 = const()[name = string("const_75_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207925760)))];
+            tensor<fp16, [1, 8, 512]> var_3052_cast_fp16 = mul(x = var_3049_cast_fp16_0, y = const_75_to_fp16)[name = string("op_3052_cast_fp16")];
+            tensor<int32, [4]> var_3058 = const()[name = string("op_3058"), val = tensor<int32, [4]>([1, 8, 1, 512])];
+            tensor<fp16, [1, 8, 1, 512]> q_35 = reshape(shape = var_3058, x = var_3052_cast_fp16)[name = string("q_35")];
+            tensor<fp16, [1, 8, 1, 512]> var_3060 = mul(x = q_35, y = cos)[name = string("op_3060")];
+            tensor<int32, [2]> var_3061_split_sizes_0 = const()[name = string("op_3061_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_3061_axis_0 = const()[name = string("op_3061_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 256]> var_3061_0, tensor<fp16, [1, 8, 1, 256]> var_3061_1 = split(axis = var_3061_axis_0, split_sizes = var_3061_split_sizes_0, x = q_35)[name = string("op_3061")];
+            fp16 const_76_promoted = const()[name = string("const_76_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 256]> var_3063 = mul(x = var_3061_1, y = const_76_promoted)[name = string("op_3063")];
+            int32 var_3065 = const()[name = string("op_3065"), val = int32(-1)];
+            bool var_3066_interleave_0 = const()[name = string("op_3066_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 512]> var_3066 = concat(axis = var_3065, interleave = var_3066_interleave_0, values = (var_3063, var_3061_0))[name = string("op_3066")];
+            tensor<fp16, [1, 8, 1, 512]> var_3067 = mul(x = var_3066, y = sin)[name = string("op_3067")];
+            tensor<fp16, [1, 8, 1, 512]> q_39 = add(x = var_3060, y = var_3067)[name = string("q_39")];
+            string var_3080_pad_type_0 = const()[name = string("op_3080_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3080_strides_0 = const()[name = string("op_3080_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3080_pad_0 = const()[name = string("op_3080_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3080_dilations_0 = const()[name = string("op_3080_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3080_groups_0 = const()[name = string("op_3080_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 512, 1, 1]> var_3080 = conv(dilations = var_3080_dilations_0, groups = var_3080_groups_0, pad = var_3080_pad_0, pad_type = var_3080_pad_type_0, strides = var_3080_strides_0, weight = layers_4_self_attn_k_proj_weight_palettized, x = var_2995_cast_fp16)[name = string("op_3080")];
+            tensor<int32, [4]> var_3085 = const()[name = string("op_3085"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 1]> var_3086 = reshape(shape = var_3085, x = var_3080)[name = string("op_3086")];
+            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_3108_pad_type_0 = const()[name = string("op_3108_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3108_strides_0 = const()[name = string("op_3108_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3108_pad_0 = const()[name = string("op_3108_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3108_dilations_0 = const()[name = string("op_3108_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3108_groups_0 = const()[name = string("op_3108_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 512, 1, 1]> var_3108 = conv(dilations = var_3108_dilations_0, groups = var_3108_groups_0, pad = var_3108_pad_0, pad_type = var_3108_pad_type_0, strides = var_3108_strides_0, weight = layers_4_self_attn_v_proj_weight_palettized, x = var_2995_cast_fp16)[name = string("op_3108")];
+            tensor<int32, [4]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 1]> var_3114 = reshape(shape = var_3113, x = var_3108)[name = string("op_3114")];
+            tensor<int32, [4]> var_3119 = const()[name = string("op_3119"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_3129 = const()[name = string("op_3129"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<fp16, [1, 1, 1, 512]> var_3092 = transpose(perm = var_3091, x = var_3086)[name = string("transpose_117")];
+            tensor<fp16, [1, 1, 512]> x_127 = reshape(shape = var_3129, x = var_3092)[name = string("x_127")];
+            int32 var_3135 = const()[name = string("op_3135"), val = int32(-1)];
+            fp16 const_77_promoted_to_fp16 = const()[name = string("const_77_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 512]> var_3141_cast_fp16 = mul(x = x_127, y = const_77_promoted_to_fp16)[name = string("op_3141_cast_fp16")];
+            bool input_127_interleave_0 = const()[name = string("input_127_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1024]> input_127_cast_fp16 = concat(axis = var_3135, interleave = input_127_interleave_0, values = (x_127, var_3141_cast_fp16))[name = string("input_127_cast_fp16")];
+            tensor<int32, [1]> normed_121_axes_0 = const()[name = string("normed_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3133_to_fp16 = const()[name = string("op_3133_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1024]> normed_121_cast_fp16 = layer_norm(axes = normed_121_axes_0, epsilon = var_3133_to_fp16, x = input_127_cast_fp16)[name = string("normed_121_cast_fp16")];
+            tensor<int32, [2]> var_3146_split_sizes_0 = const()[name = string("op_3146_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            int32 var_3146_axis_0 = const()[name = string("op_3146_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 512]> var_3146_cast_fp16_0, tensor<fp16, [1, 1, 512]> var_3146_cast_fp16_1 = split(axis = var_3146_axis_0, split_sizes = var_3146_split_sizes_0, x = normed_121_cast_fp16)[name = string("op_3146_cast_fp16")];
+            tensor<fp16, [512]> const_78_to_fp16 = const()[name = string("const_78_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207926848)))];
+            tensor<fp16, [1, 1, 512]> var_3149_cast_fp16 = mul(x = var_3146_cast_fp16_0, y = const_78_to_fp16)[name = string("op_3149_cast_fp16")];
+            tensor<int32, [4]> var_3155 = const()[name = string("op_3155"), val = tensor<int32, [4]>([1, 1, 1, 512])];
+            tensor<fp16, [1, 1, 1, 512]> q_37 = reshape(shape = var_3155, x = var_3149_cast_fp16)[name = string("q_37")];
+            fp16 var_3162_promoted_to_fp16 = const()[name = string("op_3162_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 512]> var_3120 = transpose(perm = var_3119, x = var_3114)[name = string("transpose_116")];
+            tensor<fp16, [1, 1, 1, 512]> var_3163_cast_fp16 = pow(x = var_3120, y = var_3162_promoted_to_fp16)[name = string("op_3163_cast_fp16")];
+            tensor<int32, [1]> var_3168_axes_0 = const()[name = string("op_3168_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_3168_keep_dims_0 = const()[name = string("op_3168_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_3168_cast_fp16 = reduce_mean(axes = var_3168_axes_0, keep_dims = var_3168_keep_dims_0, x = var_3163_cast_fp16)[name = string("op_3168_cast_fp16")];
+            fp16 var_3170_to_fp16 = const()[name = string("op_3170_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_9_cast_fp16 = add(x = var_3168_cast_fp16, y = var_3170_to_fp16)[name = string("mean_sq_9_cast_fp16")];
+            fp16 var_3177_to_fp16 = const()[name = string("op_3177_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_3178_cast_fp16 = pow(x = mean_sq_9_cast_fp16, y = var_3177_to_fp16)[name = string("op_3178_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 512]> var_3179_cast_fp16 = mul(x = var_3120, y = var_3178_cast_fp16)[name = string("op_3179_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 512]> var_3185 = mul(x = q_37, y = cos)[name = string("op_3185")];
+            tensor<int32, [2]> var_3186_split_sizes_0 = const()[name = string("op_3186_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_3186_axis_0 = const()[name = string("op_3186_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 256]> var_3186_0, tensor<fp16, [1, 1, 1, 256]> var_3186_1 = split(axis = var_3186_axis_0, split_sizes = var_3186_split_sizes_0, x = q_37)[name = string("op_3186")];
+            fp16 const_79_promoted = const()[name = string("const_79_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 256]> var_3188 = mul(x = var_3186_1, y = const_79_promoted)[name = string("op_3188")];
+            int32 var_3190 = const()[name = string("op_3190"), val = int32(-1)];
+            bool var_3191_interleave_0 = const()[name = string("op_3191_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 512]> var_3191 = concat(axis = var_3190, interleave = var_3191_interleave_0, values = (var_3188, var_3186_0))[name = string("op_3191")];
+            tensor<fp16, [1, 1, 1, 512]> var_3192 = mul(x = var_3191, y = sin)[name = string("op_3192")];
+            tensor<fp16, [1, 1, 1, 512]> k_11 = add(x = var_3185, y = var_3192)[name = string("k_11")];
+            tensor<int32, [4]> var_3197_begin_0 = const()[name = string("op_3197_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> var_3197_end_0 = const()[name = string("op_3197_end_0"), val = tensor<int32, [4]>([5, 1, 512, 512])];
+            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<bool, [4]> var_3197_squeeze_mask_0 = const()[name = string("op_3197_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_3197_cast_fp16 = slice_by_index(begin = var_3197_begin_0, end = var_3197_end_0, end_mask = var_3197_end_mask_0, squeeze_mask = var_3197_squeeze_mask_0, x = coreml_update_state_31)[name = string("op_3197_cast_fp16")];
+            tensor<int32, [1]> K_c_9_axes_0 = const()[name = string("K_c_9_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_9_cast_fp16 = expand_dims(axes = K_c_9_axes_0, x = var_3197_cast_fp16)[name = string("K_c_9_cast_fp16")];
+            tensor<int32, [4]> var_3202_begin_0 = const()[name = string("op_3202_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> var_3202_end_0 = const()[name = string("op_3202_end_0"), val = tensor<int32, [4]>([17, 1, 512, 512])];
+            tensor<bool, [4]> var_3202_end_mask_0 = const()[name = string("op_3202_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_3202_squeeze_mask_0 = const()[name = string("op_3202_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_3202_cast_fp16 = slice_by_index(begin = var_3202_begin_0, end = var_3202_end_0, end_mask = var_3202_end_mask_0, squeeze_mask = var_3202_squeeze_mask_0, x = coreml_update_state_31)[name = string("op_3202_cast_fp16")];
+            tensor<int32, [1]> V_c_9_axes_0 = const()[name = string("V_c_9_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_9_cast_fp16 = expand_dims(axes = V_c_9_axes_0, x = var_3202_cast_fp16)[name = string("V_c_9_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_3208_cast_fp16 = mul(x = K_c_9_cast_fp16, y = var_1005_cast_fp16)[name = string("op_3208_cast_fp16")];
+            tensor<int32, [4]> var_3209_reps_0 = const()[name = string("op_3209_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_3209 = tile(reps = var_3209_reps_0, x = k_11)[name = string("op_3209")];
+            tensor<fp16, [1, 1, 512, 512]> var_3210_cast_fp16 = mul(x = var_3209, y = update_mask)[name = string("op_3210_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_9_cast_fp16 = add(x = var_3208_cast_fp16, y = var_3210_cast_fp16)[name = string("K_n_9_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_3216_cast_fp16 = mul(x = V_c_9_cast_fp16, y = var_1005_cast_fp16)[name = string("op_3216_cast_fp16")];
+            tensor<int32, [4]> var_3217_reps_0 = const()[name = string("op_3217_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_3217 = tile(reps = var_3217_reps_0, x = var_3179_cast_fp16)[name = string("op_3217")];
+            tensor<fp16, [1, 1, 512, 512]> var_3218_cast_fp16 = mul(x = var_3217, y = update_mask)[name = string("op_3218_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_9_cast_fp16 = add(x = var_3216_cast_fp16, y = var_3218_cast_fp16)[name = string("V_n_9_cast_fp16")];
+            tensor<int32, [1]> var_3222_axes_0 = const()[name = string("op_3222_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_3222_cast_fp16 = squeeze(axes = var_3222_axes_0, x = K_n_9_cast_fp16)[name = string("op_3222_cast_fp16")];
+            tensor<int32, [4]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_9_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_9_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_9_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_32, begin_mask = kv_cache_0_internal_tensor_assign_9_begin_mask_0, end = concat_33, end_mask = kv_cache_0_internal_tensor_assign_9_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_9_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_9_stride_0, update = var_3222_cast_fp16, x = coreml_update_state_31)[name = string("kv_cache_0_internal_tensor_assign_9_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_9_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_32_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_32 = read_state(input = kv_cache_0)[name = string("coreml_update_state_32")];
+            tensor<int32, [1]> var_3229_axes_0 = const()[name = string("op_3229_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_3229_cast_fp16 = squeeze(axes = var_3229_axes_0, x = V_n_9_cast_fp16)[name = string("op_3229_cast_fp16")];
+            tensor<int32, [4]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_10_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_10_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_10_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_34, begin_mask = kv_cache_0_internal_tensor_assign_10_begin_mask_0, end = concat_35, end_mask = kv_cache_0_internal_tensor_assign_10_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_10_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_10_stride_0, update = var_3229_cast_fp16, x = coreml_update_state_32)[name = string("kv_cache_0_internal_tensor_assign_10_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_10_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_33_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_33 = read_state(input = kv_cache_0)[name = string("coreml_update_state_33")];
+            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 512]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = K_n_9_cast_fp16)[name = string("transpose_115")];
+            tensor<fp16, [8, 1, 512, 512]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
+            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([8, 1, 1, 512, 512])];
+            tensor<fp16, [8, 1, 1, 512, 512]> reshape_16_cast_fp16 = reshape(shape = concat_36, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
+            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
+            tensor<fp16, [1, 8, 1, 512, 512]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_114")];
+            tensor<fp16, [8, 1, 512, 512]> reshape_17_cast_fp16 = reshape(shape = concat_37, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
+            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 512]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = V_n_9_cast_fp16)[name = string("transpose_113")];
+            tensor<fp16, [8, 1, 512, 512]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
+            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([8, 1, 1, 512, 512])];
+            tensor<fp16, [8, 1, 1, 512, 512]> reshape_18_cast_fp16 = reshape(shape = concat_38, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
+            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
+            tensor<fp16, [1, 8, 1, 512, 512]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_112")];
+            tensor<fp16, [8, 1, 512, 512]> reshape_19_cast_fp16 = reshape(shape = concat_39, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
+            tensor<int32, [4]> Ve_9_perm_0 = const()[name = string("Ve_9_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_3266_transpose_x_0 = const()[name = string("op_3266_transpose_x_0"), val = bool(false)];
+            bool var_3266_transpose_y_0 = const()[name = string("op_3266_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 512]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_111")];
+            tensor<fp16, [1, 8, 1, 512]> var_3266_cast_fp16 = matmul(transpose_x = var_3266_transpose_x_0, transpose_y = var_3266_transpose_y_0, x = q_39, y = transpose_52_cast_fp16)[name = string("op_3266_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_3273_cast_fp16 = add(x = var_3266_cast_fp16, y = causal_mask)[name = string("op_3273_cast_fp16")];
+            int32 var_3274 = const()[name = string("op_3274"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_3276_cast_fp16 = softmax(axis = var_3274, x = var_3273_cast_fp16)[name = string("op_3276_cast_fp16")];
+            bool var_3292_transpose_x_0 = const()[name = string("op_3292_transpose_x_0"), val = bool(false)];
+            bool var_3292_transpose_y_0 = const()[name = string("op_3292_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 512]> Ve_9_cast_fp16 = transpose(perm = Ve_9_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_110")];
+            tensor<fp16, [1, 8, 1, 512]> var_3292_cast_fp16 = matmul(transpose_x = var_3292_transpose_x_0, transpose_y = var_3292_transpose_y_0, x = var_3276_cast_fp16, y = Ve_9_cast_fp16)[name = string("op_3292_cast_fp16")];
+            tensor<int32, [4]> var_3302 = const()[name = string("op_3302"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3309 = const()[name = string("op_3309"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 512]> var_3303 = transpose(perm = var_3302, x = var_3292_cast_fp16)[name = string("transpose_109")];
+            tensor<fp16, [1, 1, 4096]> var_3310 = reshape(shape = var_3309, x = var_3303)[name = string("op_3310")];
+            tensor<int32, [3]> var_3314 = const()[name = string("op_3314"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 4096, 1]> squeeze_4_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207927936))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211073728))))[name = string("squeeze_4_palettized")];
+            string var_3330_pad_type_0 = const()[name = string("op_3330_pad_type_0"), val = string("valid")];
+            int32 var_3330_groups_0 = const()[name = string("op_3330_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_3330_strides_0 = const()[name = string("op_3330_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_3330_pad_0 = const()[name = string("op_3330_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_3330_dilations_0 = const()[name = string("op_3330_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 4096, 1]> var_3315 = transpose(perm = var_3314, x = var_3310)[name = string("transpose_108")];
+            tensor<fp16, [1, 1536, 1]> var_3330 = conv(dilations = var_3330_dilations_0, groups = var_3330_groups_0, pad = var_3330_pad_0, pad_type = var_3330_pad_type_0, strides = var_3330_strides_0, weight = squeeze_4_palettized, x = var_3315)[name = string("op_3330")];
+            tensor<int32, [3]> var_3334 = const()[name = string("op_3334"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_3340 = const()[name = string("op_3340"), val = int32(-1)];
+            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_133 = transpose(perm = var_3334, x = var_3330)[name = string("transpose_107")];
+            tensor<fp16, [1, 1, 1536]> var_3346_cast_fp16 = mul(x = x_133, y = const_80_promoted_to_fp16)[name = string("op_3346_cast_fp16")];
+            bool input_131_interleave_0 = const()[name = string("input_131_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_131_cast_fp16 = concat(axis = var_3340, interleave = input_131_interleave_0, values = (x_133, var_3346_cast_fp16))[name = string("input_131_cast_fp16")];
+            tensor<int32, [1]> normed_125_axes_0 = const()[name = string("normed_125_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3338_to_fp16 = const()[name = string("op_3338_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_125_cast_fp16 = layer_norm(axes = normed_125_axes_0, epsilon = var_3338_to_fp16, x = input_131_cast_fp16)[name = string("normed_125_cast_fp16")];
+            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3351_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3351_cast_fp16_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = normed_125_cast_fp16)[name = string("op_3351_cast_fp16")];
+            tensor<fp16, [1536]> const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211075328)))];
+            tensor<fp16, [1, 1, 1536]> var_3354_cast_fp16 = mul(x = var_3351_cast_fp16_0, y = const_81_to_fp16)[name = string("op_3354_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_137_cast_fp16 = add(x = x_119_cast_fp16, y = var_3354_cast_fp16)[name = string("x_137_cast_fp16")];
+            int32 var_3361 = const()[name = string("op_3361"), val = int32(-1)];
+            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3367_cast_fp16 = mul(x = x_137_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3367_cast_fp16")];
+            bool input_133_interleave_0 = const()[name = string("input_133_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_133_cast_fp16 = concat(axis = var_3361, interleave = input_133_interleave_0, values = (x_137_cast_fp16, var_3367_cast_fp16))[name = string("input_133_cast_fp16")];
+            tensor<int32, [1]> normed_129_axes_0 = const()[name = string("normed_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3359_to_fp16 = const()[name = string("op_3359_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_129_cast_fp16 = layer_norm(axes = normed_129_axes_0, epsilon = var_3359_to_fp16, x = input_133_cast_fp16)[name = string("normed_129_cast_fp16")];
+            tensor<int32, [2]> var_3372_split_sizes_0 = const()[name = string("op_3372_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3372_axis_0 = const()[name = string("op_3372_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3372_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3372_cast_fp16_1 = split(axis = var_3372_axis_0, split_sizes = var_3372_split_sizes_0, x = normed_129_cast_fp16)[name = string("op_3372_cast_fp16")];
+            tensor<fp16, [1536]> const_83_to_fp16 = const()[name = string("const_83_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211078464)))];
+            tensor<fp16, [1, 1, 1536]> var_3375_cast_fp16 = mul(x = var_3372_cast_fp16_0, y = const_83_to_fp16)[name = string("op_3375_cast_fp16")];
+            tensor<int32, [3]> var_3388 = const()[name = string("op_3388"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_135_axes_0 = const()[name = string("input_135_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3389 = transpose(perm = var_3388, x = var_3375_cast_fp16)[name = string("transpose_106")];
+            tensor<fp16, [1, 1536, 1, 1]> input_135 = expand_dims(axes = input_135_axes_0, x = var_3389)[name = string("input_135")];
+            string var_3402_pad_type_0 = const()[name = string("op_3402_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3402_strides_0 = const()[name = string("op_3402_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3402_pad_0 = const()[name = string("op_3402_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3402_dilations_0 = const()[name = string("op_3402_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3402_groups_0 = const()[name = string("op_3402_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_3402 = conv(dilations = var_3402_dilations_0, groups = var_3402_groups_0, pad = var_3402_pad_0, pad_type = var_3402_pad_type_0, strides = var_3402_strides_0, weight = layers_4_mlp_gate_proj_weight_palettized, x = input_135)[name = string("op_3402")];
+            string var_3404_mode_0 = const()[name = string("op_3404_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_3404 = gelu(mode = var_3404_mode_0, x = var_3402)[name = string("op_3404")];
+            string var_3415_pad_type_0 = const()[name = string("op_3415_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3415_strides_0 = const()[name = string("op_3415_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3415_pad_0 = const()[name = string("op_3415_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3415_dilations_0 = const()[name = string("op_3415_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3415_groups_0 = const()[name = string("op_3415_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_3415 = conv(dilations = var_3415_dilations_0, groups = var_3415_groups_0, pad = var_3415_pad_0, pad_type = var_3415_pad_type_0, strides = var_3415_strides_0, weight = layers_4_mlp_up_proj_weight_palettized, x = input_135)[name = string("op_3415")];
+            tensor<fp16, [1, 6144, 1, 1]> input_137 = mul(x = var_3404, y = var_3415)[name = string("input_137")];
+            string var_3427_pad_type_0 = const()[name = string("op_3427_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3427_strides_0 = const()[name = string("op_3427_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3427_pad_0 = const()[name = string("op_3427_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3427_dilations_0 = const()[name = string("op_3427_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3427_groups_0 = const()[name = string("op_3427_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_3427 = conv(dilations = var_3427_dilations_0, groups = var_3427_groups_0, pad = var_3427_pad_0, pad_type = var_3427_pad_type_0, strides = var_3427_strides_0, weight = layers_4_mlp_down_proj_weight_palettized, x = input_137)[name = string("op_3427")];
+            tensor<int32, [1]> var_3429_axes_0 = const()[name = string("op_3429_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3429 = squeeze(axes = var_3429_axes_0, x = var_3427)[name = string("op_3429")];
+            tensor<int32, [3]> var_3433 = const()[name = string("op_3433"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_3439 = const()[name = string("op_3439"), 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]> x_141 = transpose(perm = var_3433, x = var_3429)[name = string("transpose_105")];
+            tensor<fp16, [1, 1, 1536]> var_3445_cast_fp16 = mul(x = x_141, y = const_84_promoted_to_fp16)[name = string("op_3445_cast_fp16")];
+            bool input_139_interleave_0 = const()[name = string("input_139_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_139_cast_fp16 = concat(axis = var_3439, interleave = input_139_interleave_0, values = (x_141, var_3445_cast_fp16))[name = string("input_139_cast_fp16")];
+            tensor<int32, [1]> normed_133_axes_0 = const()[name = string("normed_133_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_133_cast_fp16 = layer_norm(axes = normed_133_axes_0, epsilon = var_3437_to_fp16, x = input_139_cast_fp16)[name = string("normed_133_cast_fp16")];
+            tensor<int32, [2]> var_3450_split_sizes_0 = const()[name = string("op_3450_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3450_axis_0 = const()[name = string("op_3450_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3450_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3450_cast_fp16_1 = split(axis = var_3450_axis_0, split_sizes = var_3450_split_sizes_0, x = normed_133_cast_fp16)[name = string("op_3450_cast_fp16")];
+            tensor<fp16, [1536]> const_85_to_fp16 = const()[name = string("const_85_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211081600)))];
+            tensor<fp16, [1, 1, 1536]> var_3453_cast_fp16 = mul(x = var_3450_cast_fp16_0, y = const_85_to_fp16)[name = string("op_3453_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_69_cast_fp16 = add(x = x_137_cast_fp16, y = var_3453_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_3464 = linear(bias = linear_0_bias_0, weight = layers_4_per_layer_input_gate_weight_palettized, x = hidden_states_69_cast_fp16)[name = string("linear_8")];
+            string gated_9_mode_0 = const()[name = string("gated_9_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_9 = gelu(mode = gated_9_mode_0, x = var_3464)[name = string("gated_9")];
+            tensor<int32, [3]> var_3481_begin_0 = const()[name = string("op_3481_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
+            tensor<int32, [3]> var_3481_end_0 = const()[name = string("op_3481_end_0"), val = tensor<int32, [3]>([1, 1, 1280])];
+            tensor<bool, [3]> var_3481_end_mask_0 = const()[name = string("op_3481_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_3481_cast_fp16 = slice_by_index(begin = var_3481_begin_0, end = var_3481_end_0, end_mask = var_3481_end_mask_0, x = per_layer_combined)[name = string("op_3481_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_143_cast_fp16 = mul(x = gated_9, y = var_3481_cast_fp16)[name = string("input_143_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_4_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211084736))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211281408))))[name = string("layers_4_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_143_cast_fp16)[name = string("linear_9_cast_fp16")];
+            int32 var_3490 = const()[name = string("op_3490"), val = int32(-1)];
+            fp16 const_86_promoted_to_fp16 = const()[name = string("const_86_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3496_cast_fp16 = mul(x = linear_9_cast_fp16, y = const_86_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
+            bool input_145_interleave_0 = const()[name = string("input_145_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_145_cast_fp16 = concat(axis = var_3490, interleave = input_145_interleave_0, values = (linear_9_cast_fp16, var_3496_cast_fp16))[name = string("input_145_cast_fp16")];
+            tensor<int32, [1]> normed_137_axes_0 = const()[name = string("normed_137_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3488_to_fp16 = const()[name = string("op_3488_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_137_cast_fp16 = layer_norm(axes = normed_137_axes_0, epsilon = var_3488_to_fp16, x = input_145_cast_fp16)[name = string("normed_137_cast_fp16")];
+            tensor<int32, [2]> var_3501_split_sizes_0 = const()[name = string("op_3501_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3501_axis_0 = const()[name = string("op_3501_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3501_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3501_cast_fp16_1 = split(axis = var_3501_axis_0, split_sizes = var_3501_split_sizes_0, x = normed_137_cast_fp16)[name = string("op_3501_cast_fp16")];
+            tensor<fp16, [1536]> const_87_to_fp16 = const()[name = string("const_87_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211283008)))];
+            tensor<fp16, [1, 1, 1536]> var_3504_cast_fp16 = mul(x = var_3501_cast_fp16_0, y = const_87_to_fp16)[name = string("op_3504_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_73_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = var_3504_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
+            tensor<fp16, [1]> layers_4_layer_scalar_to_fp16 = const()[name = string("layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fep-2])];
+            tensor<fp16, [1, 1, 1536]> x_149_cast_fp16 = mul(x = hidden_states_73_cast_fp16, y = layers_4_layer_scalar_to_fp16)[name = string("x_149_cast_fp16")];
+            int32 var_3512 = const()[name = string("op_3512"), 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_3518_cast_fp16 = mul(x = x_149_cast_fp16, y = const_88_promoted_to_fp16)[name = string("op_3518_cast_fp16")];
+            bool input_147_interleave_0 = const()[name = string("input_147_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_147_cast_fp16 = concat(axis = var_3512, interleave = input_147_interleave_0, values = (x_149_cast_fp16, var_3518_cast_fp16))[name = string("input_147_cast_fp16")];
+            tensor<int32, [1]> normed_141_axes_0 = const()[name = string("normed_141_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3510_to_fp16 = const()[name = string("op_3510_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_141_cast_fp16 = layer_norm(axes = normed_141_axes_0, epsilon = var_3510_to_fp16, x = input_147_cast_fp16)[name = string("normed_141_cast_fp16")];
+            tensor<int32, [2]> var_3523_split_sizes_0 = const()[name = string("op_3523_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3523_axis_0 = const()[name = string("op_3523_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3523_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3523_cast_fp16_1 = split(axis = var_3523_axis_0, split_sizes = var_3523_split_sizes_0, x = normed_141_cast_fp16)[name = string("op_3523_cast_fp16")];
+            tensor<fp16, [1536]> const_89_to_fp16 = const()[name = string("const_89_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211286144)))];
+            tensor<fp16, [1, 1, 1536]> var_3526_cast_fp16 = mul(x = var_3523_cast_fp16_0, y = const_89_to_fp16)[name = string("op_3526_cast_fp16")];
+            tensor<int32, [3]> var_3534 = const()[name = string("op_3534"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_3537_axes_0 = const()[name = string("op_3537_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3535_cast_fp16 = transpose(perm = var_3534, x = var_3526_cast_fp16)[name = string("transpose_104")];
+            tensor<fp16, [1, 1536, 1, 1]> var_3537_cast_fp16 = expand_dims(axes = var_3537_axes_0, x = var_3535_cast_fp16)[name = string("op_3537_cast_fp16")];
+            string var_3553_pad_type_0 = const()[name = string("op_3553_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3553_strides_0 = const()[name = string("op_3553_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3553_pad_0 = const()[name = string("op_3553_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3553_dilations_0 = const()[name = string("op_3553_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3553_groups_0 = const()[name = string("op_3553_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_3553 = conv(dilations = var_3553_dilations_0, groups = var_3553_groups_0, pad = var_3553_pad_0, pad_type = var_3553_pad_type_0, strides = var_3553_strides_0, weight = layers_5_self_attn_q_proj_weight_palettized, x = var_3537_cast_fp16)[name = string("op_3553")];
+            tensor<int32, [4]> var_3558 = const()[name = string("op_3558"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_3559 = reshape(shape = var_3558, x = var_3553)[name = string("op_3559")];
+            tensor<int32, [4]> var_3564 = const()[name = string("op_3564"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_3574 = const()[name = string("op_3574"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_3565 = transpose(perm = var_3564, x = var_3559)[name = string("transpose_103")];
+            tensor<fp16, [1, 8, 256]> x_153 = reshape(shape = var_3574, x = var_3565)[name = string("x_153")];
+            int32 var_3580 = const()[name = string("op_3580"), val = int32(-1)];
+            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_3586_cast_fp16 = mul(x = x_153, y = const_90_promoted_to_fp16)[name = string("op_3586_cast_fp16")];
+            bool input_151_interleave_0 = const()[name = string("input_151_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_151_cast_fp16 = concat(axis = var_3580, interleave = input_151_interleave_0, values = (x_153, var_3586_cast_fp16))[name = string("input_151_cast_fp16")];
+            tensor<int32, [1]> normed_145_axes_0 = const()[name = string("normed_145_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3578_to_fp16 = const()[name = string("op_3578_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_145_cast_fp16 = layer_norm(axes = normed_145_axes_0, epsilon = var_3578_to_fp16, x = input_151_cast_fp16)[name = string("normed_145_cast_fp16")];
+            tensor<int32, [2]> var_3591_split_sizes_0 = const()[name = string("op_3591_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_3591_axis_0 = const()[name = string("op_3591_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_3591_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_3591_cast_fp16_1 = split(axis = var_3591_axis_0, split_sizes = var_3591_split_sizes_0, x = normed_145_cast_fp16)[name = string("op_3591_cast_fp16")];
+            tensor<fp16, [256]> const_91_to_fp16 = const()[name = string("const_91_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211289280)))];
+            tensor<fp16, [1, 8, 256]> var_3594_cast_fp16 = mul(x = var_3591_cast_fp16_0, y = const_91_to_fp16)[name = string("op_3594_cast_fp16")];
+            tensor<int32, [4]> var_3600 = const()[name = string("op_3600"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_43 = reshape(shape = var_3600, x = var_3594_cast_fp16)[name = string("q_43")];
+            tensor<fp16, [1, 8, 1, 256]> var_3602 = mul(x = q_43, y = cos_1)[name = string("op_3602")];
+            tensor<int32, [2]> var_3603_split_sizes_0 = const()[name = string("op_3603_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_3603_axis_0 = const()[name = string("op_3603_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_3603_0, tensor<fp16, [1, 8, 1, 128]> var_3603_1 = split(axis = var_3603_axis_0, split_sizes = var_3603_split_sizes_0, x = q_43)[name = string("op_3603")];
+            fp16 const_92_promoted = const()[name = string("const_92_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_3605 = mul(x = var_3603_1, y = const_92_promoted)[name = string("op_3605")];
+            int32 var_3607 = const()[name = string("op_3607"), val = int32(-1)];
+            bool var_3608_interleave_0 = const()[name = string("op_3608_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_3608 = concat(axis = var_3607, interleave = var_3608_interleave_0, values = (var_3605, var_3603_0))[name = string("op_3608")];
+            tensor<fp16, [1, 8, 1, 256]> var_3609 = mul(x = var_3608, y = sin_1)[name = string("op_3609")];
+            tensor<fp16, [1, 8, 1, 256]> q_47 = add(x = var_3602, y = var_3609)[name = string("q_47")];
+            string var_3622_pad_type_0 = const()[name = string("op_3622_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3622_strides_0 = const()[name = string("op_3622_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3622_pad_0 = const()[name = string("op_3622_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3622_dilations_0 = const()[name = string("op_3622_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3622_groups_0 = const()[name = string("op_3622_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_3622 = conv(dilations = var_3622_dilations_0, groups = var_3622_groups_0, pad = var_3622_pad_0, pad_type = var_3622_pad_type_0, strides = var_3622_strides_0, weight = layers_5_self_attn_k_proj_weight_palettized, x = var_3537_cast_fp16)[name = string("op_3622")];
+            tensor<int32, [4]> var_3627 = const()[name = string("op_3627"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_3628 = reshape(shape = var_3627, x = var_3622)[name = string("op_3628")];
+            tensor<int32, [4]> var_3633 = const()[name = string("op_3633"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_3650_pad_type_0 = const()[name = string("op_3650_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3650_strides_0 = const()[name = string("op_3650_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3650_pad_0 = const()[name = string("op_3650_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3650_dilations_0 = const()[name = string("op_3650_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3650_groups_0 = const()[name = string("op_3650_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_3650 = conv(dilations = var_3650_dilations_0, groups = var_3650_groups_0, pad = var_3650_pad_0, pad_type = var_3650_pad_type_0, strides = var_3650_strides_0, weight = layers_5_self_attn_v_proj_weight_palettized, x = var_3537_cast_fp16)[name = string("op_3650")];
+            tensor<int32, [4]> var_3655 = const()[name = string("op_3655"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_3656 = reshape(shape = var_3655, x = var_3650)[name = string("op_3656")];
+            tensor<int32, [4]> var_3661 = const()[name = string("op_3661"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_3671 = const()[name = string("op_3671"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_3634 = transpose(perm = var_3633, x = var_3628)[name = string("transpose_102")];
+            tensor<fp16, [1, 1, 256]> x_157 = reshape(shape = var_3671, x = var_3634)[name = string("x_157")];
+            int32 var_3677 = const()[name = string("op_3677"), val = int32(-1)];
+            fp16 const_93_promoted_to_fp16 = const()[name = string("const_93_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_3683_cast_fp16 = mul(x = x_157, y = const_93_promoted_to_fp16)[name = string("op_3683_cast_fp16")];
+            bool input_153_interleave_0 = const()[name = string("input_153_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_153_cast_fp16 = concat(axis = var_3677, interleave = input_153_interleave_0, values = (x_157, var_3683_cast_fp16))[name = string("input_153_cast_fp16")];
+            tensor<int32, [1]> normed_149_axes_0 = const()[name = string("normed_149_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3675_to_fp16 = const()[name = string("op_3675_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_149_cast_fp16 = layer_norm(axes = normed_149_axes_0, epsilon = var_3675_to_fp16, x = input_153_cast_fp16)[name = string("normed_149_cast_fp16")];
+            tensor<int32, [2]> var_3688_split_sizes_0 = const()[name = string("op_3688_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_3688_axis_0 = const()[name = string("op_3688_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_3688_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_3688_cast_fp16_1 = split(axis = var_3688_axis_0, split_sizes = var_3688_split_sizes_0, x = normed_149_cast_fp16)[name = string("op_3688_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_3691_cast_fp16 = mul(x = var_3688_cast_fp16_0, y = const_6_to_fp16)[name = string("op_3691_cast_fp16")];
+            tensor<int32, [4]> var_3697 = const()[name = string("op_3697"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_45 = reshape(shape = var_3697, x = var_3691_cast_fp16)[name = string("q_45")];
+            fp16 var_3704_promoted_to_fp16 = const()[name = string("op_3704_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_3662 = transpose(perm = var_3661, x = var_3656)[name = string("transpose_101")];
+            tensor<fp16, [1, 1, 1, 256]> var_3705_cast_fp16 = pow(x = var_3662, y = var_3704_promoted_to_fp16)[name = string("op_3705_cast_fp16")];
+            tensor<int32, [1]> var_3710_axes_0 = const()[name = string("op_3710_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_3710_keep_dims_0 = const()[name = string("op_3710_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_3710_cast_fp16 = reduce_mean(axes = var_3710_axes_0, keep_dims = var_3710_keep_dims_0, x = var_3705_cast_fp16)[name = string("op_3710_cast_fp16")];
+            fp16 var_3712_to_fp16 = const()[name = string("op_3712_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_11_cast_fp16 = add(x = var_3710_cast_fp16, y = var_3712_to_fp16)[name = string("mean_sq_11_cast_fp16")];
+            fp16 var_3719_to_fp16 = const()[name = string("op_3719_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_3720_cast_fp16 = pow(x = mean_sq_11_cast_fp16, y = var_3719_to_fp16)[name = string("op_3720_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_3721_cast_fp16 = mul(x = var_3662, y = var_3720_cast_fp16)[name = string("op_3721_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_3727 = mul(x = q_45, y = cos_1)[name = string("op_3727")];
+            tensor<int32, [2]> var_3728_split_sizes_0 = const()[name = string("op_3728_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_3728_axis_0 = const()[name = string("op_3728_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_3728_0, tensor<fp16, [1, 1, 1, 128]> var_3728_1 = split(axis = var_3728_axis_0, split_sizes = var_3728_split_sizes_0, x = q_45)[name = string("op_3728")];
+            fp16 const_95_promoted = const()[name = string("const_95_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_3730 = mul(x = var_3728_1, y = const_95_promoted)[name = string("op_3730")];
+            int32 var_3732 = const()[name = string("op_3732"), val = int32(-1)];
+            bool var_3733_interleave_0 = const()[name = string("op_3733_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_3733 = concat(axis = var_3732, interleave = var_3733_interleave_0, values = (var_3730, var_3728_0))[name = string("op_3733")];
+            tensor<fp16, [1, 1, 1, 256]> var_3734 = mul(x = var_3733, y = sin_1)[name = string("op_3734")];
+            tensor<fp16, [1, 1, 1, 256]> input_155 = add(x = var_3727, y = var_3734)[name = string("input_155")];
+            tensor<int32, [4]> var_3739_begin_0 = const()[name = string("op_3739_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> var_3739_end_0 = const()[name = string("op_3739_end_0"), val = tensor<int32, [4]>([6, 1, 512, 512])];
+            tensor<bool, [4]> var_3739_end_mask_0 = const()[name = string("op_3739_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_3739_squeeze_mask_0 = const()[name = string("op_3739_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_3739_cast_fp16 = slice_by_index(begin = var_3739_begin_0, end = var_3739_end_0, end_mask = var_3739_end_mask_0, squeeze_mask = var_3739_squeeze_mask_0, x = coreml_update_state_33)[name = string("op_3739_cast_fp16")];
+            tensor<int32, [1]> K_c_11_axes_0 = const()[name = string("K_c_11_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_11_cast_fp16 = expand_dims(axes = K_c_11_axes_0, x = var_3739_cast_fp16)[name = string("K_c_11_cast_fp16")];
+            tensor<int32, [4]> var_3744_begin_0 = const()[name = string("op_3744_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> var_3744_end_0 = const()[name = string("op_3744_end_0"), val = tensor<int32, [4]>([18, 1, 512, 512])];
+            tensor<bool, [4]> var_3744_end_mask_0 = const()[name = string("op_3744_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_3744_squeeze_mask_0 = const()[name = string("op_3744_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_3744_cast_fp16 = slice_by_index(begin = var_3744_begin_0, end = var_3744_end_0, end_mask = var_3744_end_mask_0, squeeze_mask = var_3744_squeeze_mask_0, x = coreml_update_state_33)[name = string("op_3744_cast_fp16")];
+            tensor<int32, [1]> V_c_11_axes_0 = const()[name = string("V_c_11_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_11_cast_fp16 = expand_dims(axes = V_c_11_axes_0, x = var_3744_cast_fp16)[name = string("V_c_11_cast_fp16")];
+            tensor<int32, [8]> kp_9_pad_0 = const()[name = string("kp_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_9_mode_0 = const()[name = string("kp_9_mode_0"), val = string("constant")];
+            fp16 const_96_to_fp16 = const()[name = string("const_96_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_9_cast_fp16 = pad(constant_val = const_96_to_fp16, mode = kp_9_mode_0, pad = kp_9_pad_0, x = input_155)[name = string("kp_9_cast_fp16")];
+            tensor<int32, [8]> vp_9_pad_0 = const()[name = string("vp_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_9_mode_0 = const()[name = string("vp_9_mode_0"), val = string("constant")];
+            fp16 const_97_to_fp16 = const()[name = string("const_97_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_9_cast_fp16 = pad(constant_val = const_97_to_fp16, mode = vp_9_mode_0, pad = vp_9_pad_0, x = var_3721_cast_fp16)[name = string("vp_9_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_3762_cast_fp16 = mul(x = K_c_11_cast_fp16, y = var_1005_cast_fp16)[name = string("op_3762_cast_fp16")];
+            tensor<int32, [4]> var_3763_reps_0 = const()[name = string("op_3763_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_3763_cast_fp16 = tile(reps = var_3763_reps_0, x = kp_9_cast_fp16)[name = string("op_3763_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_3764_cast_fp16 = mul(x = var_3763_cast_fp16, y = update_mask)[name = string("op_3764_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_11_cast_fp16 = add(x = var_3762_cast_fp16, y = var_3764_cast_fp16)[name = string("K_n_11_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_3770_cast_fp16 = mul(x = V_c_11_cast_fp16, y = var_1005_cast_fp16)[name = string("op_3770_cast_fp16")];
+            tensor<int32, [4]> var_3771_reps_0 = const()[name = string("op_3771_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_3771_cast_fp16 = tile(reps = var_3771_reps_0, x = vp_9_cast_fp16)[name = string("op_3771_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_3772_cast_fp16 = mul(x = var_3771_cast_fp16, y = update_mask)[name = string("op_3772_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_11_cast_fp16 = add(x = var_3770_cast_fp16, y = var_3772_cast_fp16)[name = string("V_n_11_cast_fp16")];
+            tensor<int32, [1]> var_3776_axes_0 = const()[name = string("op_3776_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_3776_cast_fp16 = squeeze(axes = var_3776_axes_0, x = K_n_11_cast_fp16)[name = string("op_3776_cast_fp16")];
+            tensor<int32, [4]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_11_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_11_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_11_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_40, begin_mask = kv_cache_0_internal_tensor_assign_11_begin_mask_0, end = concat_41, end_mask = kv_cache_0_internal_tensor_assign_11_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_11_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_11_stride_0, update = var_3776_cast_fp16, x = coreml_update_state_33)[name = string("kv_cache_0_internal_tensor_assign_11_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_11_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_34_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_34 = read_state(input = kv_cache_0)[name = string("coreml_update_state_34")];
+            tensor<int32, [1]> var_3783_axes_0 = const()[name = string("op_3783_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_3783_cast_fp16 = squeeze(axes = var_3783_axes_0, x = V_n_11_cast_fp16)[name = string("op_3783_cast_fp16")];
+            tensor<int32, [4]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_12_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_12_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_12_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_42, begin_mask = kv_cache_0_internal_tensor_assign_12_begin_mask_0, end = concat_43, end_mask = kv_cache_0_internal_tensor_assign_12_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_12_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_12_stride_0, update = var_3783_cast_fp16, x = coreml_update_state_34)[name = string("kv_cache_0_internal_tensor_assign_12_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_12_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_35_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_35 = read_state(input = kv_cache_0)[name = string("coreml_update_state_35")];
+            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, x = K_n_11_cast_fp16)[name = string("op_3793_cast_fp16")];
+            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = var_3793_cast_fp16)[name = string("transpose_100")];
+            tensor<fp16, [8, 1, 512, 256]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
+            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_20_cast_fp16 = reshape(shape = concat_44, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
+            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_99")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_21_cast_fp16 = reshape(shape = concat_45, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
+            tensor<int32, [4]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_3802_begin_0 = const()[name = string("op_3802_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_3802_end_0 = const()[name = string("op_3802_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_3802_end_mask_0 = const()[name = string("op_3802_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_3802_cast_fp16 = slice_by_index(begin = var_3802_begin_0, end = var_3802_end_0, end_mask = var_3802_end_mask_0, x = V_n_11_cast_fp16)[name = string("op_3802_cast_fp16")];
+            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = var_3802_cast_fp16)[name = string("transpose_98")];
+            tensor<fp16, [8, 1, 512, 256]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
+            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_22_cast_fp16 = reshape(shape = concat_46, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
+            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_97")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_23_cast_fp16 = reshape(shape = concat_47, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
+            tensor<int32, [4]> Ve_11_perm_0 = const()[name = string("Ve_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_3820_transpose_x_0 = const()[name = string("op_3820_transpose_x_0"), val = bool(false)];
+            bool var_3820_transpose_y_0 = const()[name = string("op_3820_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_96")];
+            tensor<fp16, [1, 8, 1, 512]> var_3820_cast_fp16 = matmul(transpose_x = var_3820_transpose_x_0, transpose_y = var_3820_transpose_y_0, x = q_47, y = transpose_53_cast_fp16)[name = string("op_3820_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_3827_cast_fp16 = add(x = var_3820_cast_fp16, y = causal_mask)[name = string("op_3827_cast_fp16")];
+            int32 var_3828 = const()[name = string("op_3828"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_3830_cast_fp16 = softmax(axis = var_3828, x = var_3827_cast_fp16)[name = string("op_3830_cast_fp16")];
+            bool var_3846_transpose_x_0 = const()[name = string("op_3846_transpose_x_0"), val = bool(false)];
+            bool var_3846_transpose_y_0 = const()[name = string("op_3846_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_11_cast_fp16 = transpose(perm = Ve_11_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_95")];
+            tensor<fp16, [1, 8, 1, 256]> var_3846_cast_fp16 = matmul(transpose_x = var_3846_transpose_x_0, transpose_y = var_3846_transpose_y_0, x = var_3830_cast_fp16, y = Ve_11_cast_fp16)[name = string("op_3846_cast_fp16")];
+            tensor<int32, [4]> var_3856 = const()[name = string("op_3856"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3863 = const()[name = string("op_3863"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_3857 = transpose(perm = var_3856, x = var_3846_cast_fp16)[name = string("transpose_94")];
+            tensor<fp16, [1, 1, 2048]> var_3864 = reshape(shape = var_3863, x = var_3857)[name = string("op_3864")];
+            tensor<int32, [3]> var_3868 = const()[name = string("op_3868"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_5_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211289856))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212862784))))[name = string("squeeze_5_palettized")];
+            string var_3884_pad_type_0 = const()[name = string("op_3884_pad_type_0"), val = string("valid")];
+            int32 var_3884_groups_0 = const()[name = string("op_3884_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_3884_strides_0 = const()[name = string("op_3884_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_3884_pad_0 = const()[name = string("op_3884_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_3884_dilations_0 = const()[name = string("op_3884_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_3869 = transpose(perm = var_3868, x = var_3864)[name = string("transpose_93")];
+            tensor<fp16, [1, 1536, 1]> var_3884 = conv(dilations = var_3884_dilations_0, groups = var_3884_groups_0, pad = var_3884_pad_0, pad_type = var_3884_pad_type_0, strides = var_3884_strides_0, weight = squeeze_5_palettized, x = var_3869)[name = string("op_3884")];
+            tensor<int32, [3]> var_3888 = const()[name = string("op_3888"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_3894 = const()[name = string("op_3894"), val = int32(-1)];
+            fp16 const_98_promoted_to_fp16 = const()[name = string("const_98_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_163 = transpose(perm = var_3888, x = var_3884)[name = string("transpose_92")];
+            tensor<fp16, [1, 1, 1536]> var_3900_cast_fp16 = mul(x = x_163, y = const_98_promoted_to_fp16)[name = string("op_3900_cast_fp16")];
+            bool input_161_interleave_0 = const()[name = string("input_161_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_161_cast_fp16 = concat(axis = var_3894, interleave = input_161_interleave_0, values = (x_163, var_3900_cast_fp16))[name = string("input_161_cast_fp16")];
+            tensor<int32, [1]> normed_153_axes_0 = const()[name = string("normed_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3892_to_fp16 = const()[name = string("op_3892_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_153_cast_fp16 = layer_norm(axes = normed_153_axes_0, epsilon = var_3892_to_fp16, x = input_161_cast_fp16)[name = string("normed_153_cast_fp16")];
+            tensor<int32, [2]> var_3905_split_sizes_0 = const()[name = string("op_3905_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3905_axis_0 = const()[name = string("op_3905_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3905_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3905_cast_fp16_1 = split(axis = var_3905_axis_0, split_sizes = var_3905_split_sizes_0, x = normed_153_cast_fp16)[name = string("op_3905_cast_fp16")];
+            tensor<fp16, [1536]> const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212864384)))];
+            tensor<fp16, [1, 1, 1536]> var_3908_cast_fp16 = mul(x = var_3905_cast_fp16_0, y = const_99_to_fp16)[name = string("op_3908_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_167_cast_fp16 = add(x = x_149_cast_fp16, y = var_3908_cast_fp16)[name = string("x_167_cast_fp16")];
+            int32 var_3915 = const()[name = string("op_3915"), val = int32(-1)];
+            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_3921_cast_fp16 = mul(x = x_167_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3921_cast_fp16")];
+            bool input_163_interleave_0 = const()[name = string("input_163_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_163_cast_fp16 = concat(axis = var_3915, interleave = input_163_interleave_0, values = (x_167_cast_fp16, var_3921_cast_fp16))[name = string("input_163_cast_fp16")];
+            tensor<int32, [1]> normed_157_axes_0 = const()[name = string("normed_157_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3913_to_fp16 = const()[name = string("op_3913_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_157_cast_fp16 = layer_norm(axes = normed_157_axes_0, epsilon = var_3913_to_fp16, x = input_163_cast_fp16)[name = string("normed_157_cast_fp16")];
+            tensor<int32, [2]> var_3926_split_sizes_0 = const()[name = string("op_3926_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_3926_axis_0 = const()[name = string("op_3926_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_3926_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_3926_cast_fp16_1 = split(axis = var_3926_axis_0, split_sizes = var_3926_split_sizes_0, x = normed_157_cast_fp16)[name = string("op_3926_cast_fp16")];
+            tensor<fp16, [1536]> const_101_to_fp16 = const()[name = string("const_101_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212867520)))];
+            tensor<fp16, [1, 1, 1536]> var_3929_cast_fp16 = mul(x = var_3926_cast_fp16_0, y = const_101_to_fp16)[name = string("op_3929_cast_fp16")];
+            tensor<int32, [3]> var_3942 = const()[name = string("op_3942"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_165_axes_0 = const()[name = string("input_165_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3943 = transpose(perm = var_3942, x = var_3929_cast_fp16)[name = string("transpose_91")];
+            tensor<fp16, [1, 1536, 1, 1]> input_165 = expand_dims(axes = input_165_axes_0, x = var_3943)[name = string("input_165")];
+            string var_3956_pad_type_0 = const()[name = string("op_3956_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3956_strides_0 = const()[name = string("op_3956_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3956_pad_0 = const()[name = string("op_3956_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3956_dilations_0 = const()[name = string("op_3956_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3956_groups_0 = const()[name = string("op_3956_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_3956 = conv(dilations = var_3956_dilations_0, groups = var_3956_groups_0, pad = var_3956_pad_0, pad_type = var_3956_pad_type_0, strides = var_3956_strides_0, weight = layers_5_mlp_gate_proj_weight_palettized, x = input_165)[name = string("op_3956")];
+            string var_3958_mode_0 = const()[name = string("op_3958_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_3958 = gelu(mode = var_3958_mode_0, x = var_3956)[name = string("op_3958")];
+            string var_3969_pad_type_0 = const()[name = string("op_3969_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3969_strides_0 = const()[name = string("op_3969_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3969_pad_0 = const()[name = string("op_3969_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3969_dilations_0 = const()[name = string("op_3969_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3969_groups_0 = const()[name = string("op_3969_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_3969 = conv(dilations = var_3969_dilations_0, groups = var_3969_groups_0, pad = var_3969_pad_0, pad_type = var_3969_pad_type_0, strides = var_3969_strides_0, weight = layers_5_mlp_up_proj_weight_palettized, x = input_165)[name = string("op_3969")];
+            tensor<fp16, [1, 6144, 1, 1]> input_167 = mul(x = var_3958, y = var_3969)[name = string("input_167")];
+            string var_3981_pad_type_0 = const()[name = string("op_3981_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_3981_strides_0 = const()[name = string("op_3981_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_3981_pad_0 = const()[name = string("op_3981_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_3981_dilations_0 = const()[name = string("op_3981_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_3981_groups_0 = const()[name = string("op_3981_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_3981 = conv(dilations = var_3981_dilations_0, groups = var_3981_groups_0, pad = var_3981_pad_0, pad_type = var_3981_pad_type_0, strides = var_3981_strides_0, weight = layers_5_mlp_down_proj_weight_palettized, x = input_167)[name = string("op_3981")];
+            tensor<int32, [1]> var_3983_axes_0 = const()[name = string("op_3983_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_3983 = squeeze(axes = var_3983_axes_0, x = var_3981)[name = string("op_3983")];
+            tensor<int32, [3]> var_3987 = const()[name = string("op_3987"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_3993 = const()[name = string("op_3993"), val = int32(-1)];
+            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_171 = transpose(perm = var_3987, x = var_3983)[name = string("transpose_90")];
+            tensor<fp16, [1, 1, 1536]> var_3999_cast_fp16 = mul(x = x_171, y = const_102_promoted_to_fp16)[name = string("op_3999_cast_fp16")];
+            bool input_169_interleave_0 = const()[name = string("input_169_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_169_cast_fp16 = concat(axis = var_3993, interleave = input_169_interleave_0, values = (x_171, var_3999_cast_fp16))[name = string("input_169_cast_fp16")];
+            tensor<int32, [1]> normed_161_axes_0 = const()[name = string("normed_161_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_3991_to_fp16 = const()[name = string("op_3991_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_161_cast_fp16 = layer_norm(axes = normed_161_axes_0, epsilon = var_3991_to_fp16, x = input_169_cast_fp16)[name = string("normed_161_cast_fp16")];
+            tensor<int32, [2]> var_4004_split_sizes_0 = const()[name = string("op_4004_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4004_axis_0 = const()[name = string("op_4004_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4004_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4004_cast_fp16_1 = split(axis = var_4004_axis_0, split_sizes = var_4004_split_sizes_0, x = normed_161_cast_fp16)[name = string("op_4004_cast_fp16")];
+            tensor<fp16, [1536]> const_103_to_fp16 = const()[name = string("const_103_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212870656)))];
+            tensor<fp16, [1, 1, 1536]> var_4007_cast_fp16 = mul(x = var_4004_cast_fp16_0, y = const_103_to_fp16)[name = string("op_4007_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_83_cast_fp16 = add(x = x_167_cast_fp16, y = var_4007_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_4018 = linear(bias = linear_0_bias_0, weight = layers_5_per_layer_input_gate_weight_palettized, x = hidden_states_83_cast_fp16)[name = string("linear_10")];
+            string gated_11_mode_0 = const()[name = string("gated_11_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_11 = gelu(mode = gated_11_mode_0, x = var_4018)[name = string("gated_11")];
+            tensor<int32, [3]> var_4035_begin_0 = const()[name = string("op_4035_begin_0"), val = tensor<int32, [3]>([0, 0, 1280])];
+            tensor<int32, [3]> var_4035_end_0 = const()[name = string("op_4035_end_0"), val = tensor<int32, [3]>([1, 1, 1536])];
+            tensor<bool, [3]> var_4035_end_mask_0 = const()[name = string("op_4035_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_4035_cast_fp16 = slice_by_index(begin = var_4035_begin_0, end = var_4035_end_0, end_mask = var_4035_end_mask_0, x = per_layer_combined)[name = string("op_4035_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_173_cast_fp16 = mul(x = gated_11, y = var_4035_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_5_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212873792))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213070464))))[name = string("layers_5_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_11_cast_fp16")];
+            int32 var_4044 = const()[name = string("op_4044"), val = int32(-1)];
+            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4050_cast_fp16 = mul(x = linear_11_cast_fp16, y = const_104_promoted_to_fp16)[name = string("op_4050_cast_fp16")];
+            bool input_175_interleave_0 = const()[name = string("input_175_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_175_cast_fp16 = concat(axis = var_4044, interleave = input_175_interleave_0, values = (linear_11_cast_fp16, var_4050_cast_fp16))[name = string("input_175_cast_fp16")];
+            tensor<int32, [1]> normed_165_axes_0 = const()[name = string("normed_165_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4042_to_fp16 = const()[name = string("op_4042_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_165_cast_fp16 = layer_norm(axes = normed_165_axes_0, epsilon = var_4042_to_fp16, x = input_175_cast_fp16)[name = string("normed_165_cast_fp16")];
+            tensor<int32, [2]> var_4055_split_sizes_0 = const()[name = string("op_4055_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4055_axis_0 = const()[name = string("op_4055_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4055_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4055_cast_fp16_1 = split(axis = var_4055_axis_0, split_sizes = var_4055_split_sizes_0, x = normed_165_cast_fp16)[name = string("op_4055_cast_fp16")];
+            tensor<fp16, [1536]> const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213072064)))];
+            tensor<fp16, [1, 1, 1536]> var_4058_cast_fp16 = mul(x = var_4055_cast_fp16_0, y = const_105_to_fp16)[name = string("op_4058_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_87_cast_fp16 = add(x = hidden_states_83_cast_fp16, y = var_4058_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
+            tensor<fp16, [1]> layers_5_layer_scalar_to_fp16 = const()[name = string("layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.46p-1])];
+            tensor<fp16, [1, 1, 1536]> x_179_cast_fp16 = mul(x = hidden_states_87_cast_fp16, y = layers_5_layer_scalar_to_fp16)[name = string("x_179_cast_fp16")];
+            int32 var_4066 = const()[name = string("op_4066"), 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_4072_cast_fp16 = mul(x = x_179_cast_fp16, y = const_106_promoted_to_fp16)[name = string("op_4072_cast_fp16")];
+            bool input_177_interleave_0 = const()[name = string("input_177_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_177_cast_fp16 = concat(axis = var_4066, interleave = input_177_interleave_0, values = (x_179_cast_fp16, var_4072_cast_fp16))[name = string("input_177_cast_fp16")];
+            tensor<int32, [1]> normed_169_axes_0 = const()[name = string("normed_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4064_to_fp16 = const()[name = string("op_4064_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_169_cast_fp16 = layer_norm(axes = normed_169_axes_0, epsilon = var_4064_to_fp16, x = input_177_cast_fp16)[name = string("normed_169_cast_fp16")];
+            tensor<int32, [2]> var_4077_split_sizes_0 = const()[name = string("op_4077_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4077_axis_0 = const()[name = string("op_4077_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4077_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4077_cast_fp16_1 = split(axis = var_4077_axis_0, split_sizes = var_4077_split_sizes_0, x = normed_169_cast_fp16)[name = string("op_4077_cast_fp16")];
+            tensor<fp16, [1536]> const_107_to_fp16 = const()[name = string("const_107_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213075200)))];
+            tensor<fp16, [1, 1, 1536]> var_4080_cast_fp16 = mul(x = var_4077_cast_fp16_0, y = const_107_to_fp16)[name = string("op_4080_cast_fp16")];
+            tensor<int32, [3]> var_4088 = const()[name = string("op_4088"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_4091_axes_0 = const()[name = string("op_4091_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4089_cast_fp16 = transpose(perm = var_4088, x = var_4080_cast_fp16)[name = string("transpose_89")];
+            tensor<fp16, [1, 1536, 1, 1]> var_4091_cast_fp16 = expand_dims(axes = var_4091_axes_0, x = var_4089_cast_fp16)[name = string("op_4091_cast_fp16")];
+            string var_4107_pad_type_0 = const()[name = string("op_4107_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4107_strides_0 = const()[name = string("op_4107_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4107_pad_0 = const()[name = string("op_4107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4107_dilations_0 = const()[name = string("op_4107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4107_groups_0 = const()[name = string("op_4107_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_4107 = conv(dilations = var_4107_dilations_0, groups = var_4107_groups_0, pad = var_4107_pad_0, pad_type = var_4107_pad_type_0, strides = var_4107_strides_0, weight = layers_6_self_attn_q_proj_weight_palettized, x = var_4091_cast_fp16)[name = string("op_4107")];
+            tensor<int32, [4]> var_4112 = const()[name = string("op_4112"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_4113 = reshape(shape = var_4112, x = var_4107)[name = string("op_4113")];
+            tensor<int32, [4]> var_4118 = const()[name = string("op_4118"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_4128 = const()[name = string("op_4128"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_4119 = transpose(perm = var_4118, x = var_4113)[name = string("transpose_88")];
+            tensor<fp16, [1, 8, 256]> x_183 = reshape(shape = var_4128, x = var_4119)[name = string("x_183")];
+            int32 var_4134 = const()[name = string("op_4134"), val = int32(-1)];
+            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_4140_cast_fp16 = mul(x = x_183, y = const_108_promoted_to_fp16)[name = string("op_4140_cast_fp16")];
+            bool input_181_interleave_0 = const()[name = string("input_181_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_181_cast_fp16 = concat(axis = var_4134, interleave = input_181_interleave_0, values = (x_183, var_4140_cast_fp16))[name = string("input_181_cast_fp16")];
+            tensor<int32, [1]> normed_173_axes_0 = const()[name = string("normed_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4132_to_fp16 = const()[name = string("op_4132_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_173_cast_fp16 = layer_norm(axes = normed_173_axes_0, epsilon = var_4132_to_fp16, x = input_181_cast_fp16)[name = string("normed_173_cast_fp16")];
+            tensor<int32, [2]> var_4145_split_sizes_0 = const()[name = string("op_4145_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_4145_axis_0 = const()[name = string("op_4145_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_4145_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_4145_cast_fp16_1 = split(axis = var_4145_axis_0, split_sizes = var_4145_split_sizes_0, x = normed_173_cast_fp16)[name = string("op_4145_cast_fp16")];
+            tensor<fp16, [1, 8, 256]> var_4148_cast_fp16 = mul(x = var_4145_cast_fp16_0, y = const_21_to_fp16)[name = string("op_4148_cast_fp16")];
+            tensor<int32, [4]> var_4154 = const()[name = string("op_4154"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_51 = reshape(shape = var_4154, x = var_4148_cast_fp16)[name = string("q_51")];
+            tensor<fp16, [1, 8, 1, 256]> var_4156 = mul(x = q_51, y = cos_1)[name = string("op_4156")];
+            tensor<int32, [2]> var_4157_split_sizes_0 = const()[name = string("op_4157_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_4157_axis_0 = const()[name = string("op_4157_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_4157_0, tensor<fp16, [1, 8, 1, 128]> var_4157_1 = split(axis = var_4157_axis_0, split_sizes = var_4157_split_sizes_0, x = q_51)[name = string("op_4157")];
+            fp16 const_110_promoted = const()[name = string("const_110_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_4159 = mul(x = var_4157_1, y = const_110_promoted)[name = string("op_4159")];
+            int32 var_4161 = const()[name = string("op_4161"), val = int32(-1)];
+            bool var_4162_interleave_0 = const()[name = string("op_4162_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_4162 = concat(axis = var_4161, interleave = var_4162_interleave_0, values = (var_4159, var_4157_0))[name = string("op_4162")];
+            tensor<fp16, [1, 8, 1, 256]> var_4163 = mul(x = var_4162, y = sin_1)[name = string("op_4163")];
+            tensor<fp16, [1, 8, 1, 256]> q_55 = add(x = var_4156, y = var_4163)[name = string("q_55")];
+            string var_4176_pad_type_0 = const()[name = string("op_4176_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4176_strides_0 = const()[name = string("op_4176_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4176_pad_0 = const()[name = string("op_4176_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4176_dilations_0 = const()[name = string("op_4176_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4176_groups_0 = const()[name = string("op_4176_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_4176 = conv(dilations = var_4176_dilations_0, groups = var_4176_groups_0, pad = var_4176_pad_0, pad_type = var_4176_pad_type_0, strides = var_4176_strides_0, weight = layers_6_self_attn_k_proj_weight_palettized, x = var_4091_cast_fp16)[name = string("op_4176")];
+            tensor<int32, [4]> var_4181 = const()[name = string("op_4181"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_4182 = reshape(shape = var_4181, x = var_4176)[name = string("op_4182")];
+            tensor<int32, [4]> var_4187 = const()[name = string("op_4187"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_4204_pad_type_0 = const()[name = string("op_4204_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4204_strides_0 = const()[name = string("op_4204_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4204_pad_0 = const()[name = string("op_4204_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4204_dilations_0 = const()[name = string("op_4204_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4204_groups_0 = const()[name = string("op_4204_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_4204 = conv(dilations = var_4204_dilations_0, groups = var_4204_groups_0, pad = var_4204_pad_0, pad_type = var_4204_pad_type_0, strides = var_4204_strides_0, weight = layers_6_self_attn_v_proj_weight_palettized, x = var_4091_cast_fp16)[name = string("op_4204")];
+            tensor<int32, [4]> var_4209 = const()[name = string("op_4209"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_4210 = reshape(shape = var_4209, x = var_4204)[name = string("op_4210")];
+            tensor<int32, [4]> var_4215 = const()[name = string("op_4215"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_4225 = const()[name = string("op_4225"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_4188 = transpose(perm = var_4187, x = var_4182)[name = string("transpose_87")];
+            tensor<fp16, [1, 1, 256]> x_187 = reshape(shape = var_4225, x = var_4188)[name = string("x_187")];
+            int32 var_4231 = const()[name = string("op_4231"), val = int32(-1)];
+            fp16 const_111_promoted_to_fp16 = const()[name = string("const_111_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_4237_cast_fp16 = mul(x = x_187, y = const_111_promoted_to_fp16)[name = string("op_4237_cast_fp16")];
+            bool input_183_interleave_0 = const()[name = string("input_183_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_183_cast_fp16 = concat(axis = var_4231, interleave = input_183_interleave_0, values = (x_187, var_4237_cast_fp16))[name = string("input_183_cast_fp16")];
+            tensor<int32, [1]> normed_177_axes_0 = const()[name = string("normed_177_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4229_to_fp16 = const()[name = string("op_4229_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_177_cast_fp16 = layer_norm(axes = normed_177_axes_0, epsilon = var_4229_to_fp16, x = input_183_cast_fp16)[name = string("normed_177_cast_fp16")];
+            tensor<int32, [2]> var_4242_split_sizes_0 = const()[name = string("op_4242_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_4242_axis_0 = const()[name = string("op_4242_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_4242_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_4242_cast_fp16_1 = split(axis = var_4242_axis_0, split_sizes = var_4242_split_sizes_0, x = normed_177_cast_fp16)[name = string("op_4242_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_4245_cast_fp16 = mul(x = var_4242_cast_fp16_0, y = const_24_to_fp16)[name = string("op_4245_cast_fp16")];
+            tensor<int32, [4]> var_4251 = const()[name = string("op_4251"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_53 = reshape(shape = var_4251, x = var_4245_cast_fp16)[name = string("q_53")];
+            fp16 var_4258_promoted_to_fp16 = const()[name = string("op_4258_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_4216 = transpose(perm = var_4215, x = var_4210)[name = string("transpose_86")];
+            tensor<fp16, [1, 1, 1, 256]> var_4259_cast_fp16 = pow(x = var_4216, y = var_4258_promoted_to_fp16)[name = string("op_4259_cast_fp16")];
+            tensor<int32, [1]> var_4264_axes_0 = const()[name = string("op_4264_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_4264_keep_dims_0 = const()[name = string("op_4264_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_4264_cast_fp16 = reduce_mean(axes = var_4264_axes_0, keep_dims = var_4264_keep_dims_0, x = var_4259_cast_fp16)[name = string("op_4264_cast_fp16")];
+            fp16 var_4266_to_fp16 = const()[name = string("op_4266_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_13_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4266_to_fp16)[name = string("mean_sq_13_cast_fp16")];
+            fp16 var_4273_to_fp16 = const()[name = string("op_4273_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_4274_cast_fp16 = pow(x = mean_sq_13_cast_fp16, y = var_4273_to_fp16)[name = string("op_4274_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_4275_cast_fp16 = mul(x = var_4216, y = var_4274_cast_fp16)[name = string("op_4275_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_4281 = mul(x = q_53, y = cos_1)[name = string("op_4281")];
+            tensor<int32, [2]> var_4282_split_sizes_0 = const()[name = string("op_4282_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_4282_axis_0 = const()[name = string("op_4282_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_4282_0, tensor<fp16, [1, 1, 1, 128]> var_4282_1 = split(axis = var_4282_axis_0, split_sizes = var_4282_split_sizes_0, x = q_53)[name = string("op_4282")];
+            fp16 const_113_promoted = const()[name = string("const_113_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_4284 = mul(x = var_4282_1, y = const_113_promoted)[name = string("op_4284")];
+            int32 var_4286 = const()[name = string("op_4286"), val = int32(-1)];
+            bool var_4287_interleave_0 = const()[name = string("op_4287_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_4287 = concat(axis = var_4286, interleave = var_4287_interleave_0, values = (var_4284, var_4282_0))[name = string("op_4287")];
+            tensor<fp16, [1, 1, 1, 256]> var_4288 = mul(x = var_4287, y = sin_1)[name = string("op_4288")];
+            tensor<fp16, [1, 1, 1, 256]> input_185 = add(x = var_4281, y = var_4288)[name = string("input_185")];
+            tensor<int32, [4]> var_4293_begin_0 = const()[name = string("op_4293_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> var_4293_end_0 = const()[name = string("op_4293_end_0"), val = tensor<int32, [4]>([7, 1, 512, 512])];
+            tensor<bool, [4]> var_4293_end_mask_0 = const()[name = string("op_4293_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_4293_squeeze_mask_0 = const()[name = string("op_4293_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_4293_cast_fp16 = slice_by_index(begin = var_4293_begin_0, end = var_4293_end_0, end_mask = var_4293_end_mask_0, squeeze_mask = var_4293_squeeze_mask_0, x = coreml_update_state_35)[name = string("op_4293_cast_fp16")];
+            tensor<int32, [1]> K_c_13_axes_0 = const()[name = string("K_c_13_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_13_cast_fp16 = expand_dims(axes = K_c_13_axes_0, x = var_4293_cast_fp16)[name = string("K_c_13_cast_fp16")];
+            tensor<int32, [4]> var_4298_begin_0 = const()[name = string("op_4298_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> var_4298_end_0 = const()[name = string("op_4298_end_0"), val = tensor<int32, [4]>([19, 1, 512, 512])];
+            tensor<bool, [4]> var_4298_end_mask_0 = const()[name = string("op_4298_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_4298_squeeze_mask_0 = const()[name = string("op_4298_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_4298_cast_fp16 = slice_by_index(begin = var_4298_begin_0, end = var_4298_end_0, end_mask = var_4298_end_mask_0, squeeze_mask = var_4298_squeeze_mask_0, x = coreml_update_state_35)[name = string("op_4298_cast_fp16")];
+            tensor<int32, [1]> V_c_13_axes_0 = const()[name = string("V_c_13_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_13_cast_fp16 = expand_dims(axes = V_c_13_axes_0, x = var_4298_cast_fp16)[name = string("V_c_13_cast_fp16")];
+            tensor<int32, [8]> kp_11_pad_0 = const()[name = string("kp_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_11_mode_0 = const()[name = string("kp_11_mode_0"), val = string("constant")];
+            fp16 const_114_to_fp16 = const()[name = string("const_114_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_11_cast_fp16 = pad(constant_val = const_114_to_fp16, mode = kp_11_mode_0, pad = kp_11_pad_0, x = input_185)[name = string("kp_11_cast_fp16")];
+            tensor<int32, [8]> vp_11_pad_0 = const()[name = string("vp_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_11_mode_0 = const()[name = string("vp_11_mode_0"), val = string("constant")];
+            fp16 const_115_to_fp16 = const()[name = string("const_115_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_11_cast_fp16 = pad(constant_val = const_115_to_fp16, mode = vp_11_mode_0, pad = vp_11_pad_0, x = var_4275_cast_fp16)[name = string("vp_11_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4316_cast_fp16 = mul(x = K_c_13_cast_fp16, y = var_1005_cast_fp16)[name = string("op_4316_cast_fp16")];
+            tensor<int32, [4]> var_4317_reps_0 = const()[name = string("op_4317_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_4317_cast_fp16 = tile(reps = var_4317_reps_0, x = kp_11_cast_fp16)[name = string("op_4317_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4318_cast_fp16 = mul(x = var_4317_cast_fp16, y = update_mask)[name = string("op_4318_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_13_cast_fp16 = add(x = var_4316_cast_fp16, y = var_4318_cast_fp16)[name = string("K_n_13_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4324_cast_fp16 = mul(x = V_c_13_cast_fp16, y = var_1005_cast_fp16)[name = string("op_4324_cast_fp16")];
+            tensor<int32, [4]> var_4325_reps_0 = const()[name = string("op_4325_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_4325_cast_fp16 = tile(reps = var_4325_reps_0, x = vp_11_cast_fp16)[name = string("op_4325_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4326_cast_fp16 = mul(x = var_4325_cast_fp16, y = update_mask)[name = string("op_4326_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_13_cast_fp16 = add(x = var_4324_cast_fp16, y = var_4326_cast_fp16)[name = string("V_n_13_cast_fp16")];
+            tensor<int32, [1]> var_4330_axes_0 = const()[name = string("op_4330_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_4330_cast_fp16 = squeeze(axes = var_4330_axes_0, x = K_n_13_cast_fp16)[name = string("op_4330_cast_fp16")];
+            tensor<int32, [4]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_13_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_13_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_13_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_48, begin_mask = kv_cache_0_internal_tensor_assign_13_begin_mask_0, end = concat_49, end_mask = kv_cache_0_internal_tensor_assign_13_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_13_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_13_stride_0, update = var_4330_cast_fp16, x = coreml_update_state_35)[name = string("kv_cache_0_internal_tensor_assign_13_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_13_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_36_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_36 = read_state(input = kv_cache_0)[name = string("coreml_update_state_36")];
+            tensor<int32, [1]> var_4337_axes_0 = const()[name = string("op_4337_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_4337_cast_fp16 = squeeze(axes = var_4337_axes_0, x = V_n_13_cast_fp16)[name = string("op_4337_cast_fp16")];
+            tensor<int32, [4]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_14_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_14_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_14_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_50, begin_mask = kv_cache_0_internal_tensor_assign_14_begin_mask_0, end = concat_51, end_mask = kv_cache_0_internal_tensor_assign_14_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_14_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_14_stride_0, update = var_4337_cast_fp16, x = coreml_update_state_36)[name = string("kv_cache_0_internal_tensor_assign_14_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_14_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_37_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_37 = read_state(input = kv_cache_0)[name = string("coreml_update_state_37")];
+            tensor<int32, [4]> var_4347_begin_0 = const()[name = string("op_4347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4347_end_0 = const()[name = string("op_4347_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_4347_end_mask_0 = const()[name = string("op_4347_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_4347_cast_fp16 = slice_by_index(begin = var_4347_begin_0, end = var_4347_end_0, end_mask = var_4347_end_mask_0, x = K_n_13_cast_fp16)[name = string("op_4347_cast_fp16")];
+            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = var_4347_cast_fp16)[name = string("transpose_85")];
+            tensor<fp16, [8, 1, 512, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
+            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_24_cast_fp16 = reshape(shape = concat_52, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
+            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_84")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_25_cast_fp16 = reshape(shape = concat_53, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
+            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_4356_begin_0 = const()[name = string("op_4356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4356_end_0 = const()[name = string("op_4356_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_4356_end_mask_0 = const()[name = string("op_4356_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_4356_cast_fp16 = slice_by_index(begin = var_4356_begin_0, end = var_4356_end_0, end_mask = var_4356_end_mask_0, x = V_n_13_cast_fp16)[name = string("op_4356_cast_fp16")];
+            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = var_4356_cast_fp16)[name = string("transpose_83")];
+            tensor<fp16, [8, 1, 512, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
+            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_26_cast_fp16 = reshape(shape = concat_54, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
+            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_82")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_27_cast_fp16 = reshape(shape = concat_55, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
+            tensor<int32, [4]> Ve_13_perm_0 = const()[name = string("Ve_13_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_4374_transpose_x_0 = const()[name = string("op_4374_transpose_x_0"), val = bool(false)];
+            bool var_4374_transpose_y_0 = const()[name = string("op_4374_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_81")];
+            tensor<fp16, [1, 8, 1, 512]> var_4374_cast_fp16 = matmul(transpose_x = var_4374_transpose_x_0, transpose_y = var_4374_transpose_y_0, x = q_55, y = transpose_54_cast_fp16)[name = string("op_4374_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_4381_cast_fp16 = add(x = var_4374_cast_fp16, y = causal_mask)[name = string("op_4381_cast_fp16")];
+            int32 var_4382 = const()[name = string("op_4382"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_4384_cast_fp16 = softmax(axis = var_4382, x = var_4381_cast_fp16)[name = string("op_4384_cast_fp16")];
+            bool var_4400_transpose_x_0 = const()[name = string("op_4400_transpose_x_0"), val = bool(false)];
+            bool var_4400_transpose_y_0 = const()[name = string("op_4400_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_13_cast_fp16 = transpose(perm = Ve_13_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_80")];
+            tensor<fp16, [1, 8, 1, 256]> var_4400_cast_fp16 = matmul(transpose_x = var_4400_transpose_x_0, transpose_y = var_4400_transpose_y_0, x = var_4384_cast_fp16, y = Ve_13_cast_fp16)[name = string("op_4400_cast_fp16")];
+            tensor<int32, [4]> var_4410 = const()[name = string("op_4410"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4417 = const()[name = string("op_4417"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_4411 = transpose(perm = var_4410, x = var_4400_cast_fp16)[name = string("transpose_79")];
+            tensor<fp16, [1, 1, 2048]> var_4418 = reshape(shape = var_4417, x = var_4411)[name = string("op_4418")];
+            tensor<int32, [3]> var_4422 = const()[name = string("op_4422"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_6_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213078336))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214651264))))[name = string("squeeze_6_palettized")];
+            string var_4438_pad_type_0 = const()[name = string("op_4438_pad_type_0"), val = string("valid")];
+            int32 var_4438_groups_0 = const()[name = string("op_4438_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_4438_strides_0 = const()[name = string("op_4438_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_4438_pad_0 = const()[name = string("op_4438_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_4438_dilations_0 = const()[name = string("op_4438_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_4423 = transpose(perm = var_4422, x = var_4418)[name = string("transpose_78")];
+            tensor<fp16, [1, 1536, 1]> var_4438 = conv(dilations = var_4438_dilations_0, groups = var_4438_groups_0, pad = var_4438_pad_0, pad_type = var_4438_pad_type_0, strides = var_4438_strides_0, weight = squeeze_6_palettized, x = var_4423)[name = string("op_4438")];
+            tensor<int32, [3]> var_4442 = const()[name = string("op_4442"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_4448 = const()[name = string("op_4448"), val = int32(-1)];
+            fp16 const_116_promoted_to_fp16 = const()[name = string("const_116_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_193 = transpose(perm = var_4442, x = var_4438)[name = string("transpose_77")];
+            tensor<fp16, [1, 1, 1536]> var_4454_cast_fp16 = mul(x = x_193, y = const_116_promoted_to_fp16)[name = string("op_4454_cast_fp16")];
+            bool input_191_interleave_0 = const()[name = string("input_191_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_191_cast_fp16 = concat(axis = var_4448, interleave = input_191_interleave_0, values = (x_193, var_4454_cast_fp16))[name = string("input_191_cast_fp16")];
+            tensor<int32, [1]> normed_181_axes_0 = const()[name = string("normed_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4446_to_fp16 = const()[name = string("op_4446_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_181_cast_fp16 = layer_norm(axes = normed_181_axes_0, epsilon = var_4446_to_fp16, x = input_191_cast_fp16)[name = string("normed_181_cast_fp16")];
+            tensor<int32, [2]> var_4459_split_sizes_0 = const()[name = string("op_4459_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4459_axis_0 = const()[name = string("op_4459_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4459_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4459_cast_fp16_1 = split(axis = var_4459_axis_0, split_sizes = var_4459_split_sizes_0, x = normed_181_cast_fp16)[name = string("op_4459_cast_fp16")];
+            tensor<fp16, [1536]> const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214652864)))];
+            tensor<fp16, [1, 1, 1536]> var_4462_cast_fp16 = mul(x = var_4459_cast_fp16_0, y = const_117_to_fp16)[name = string("op_4462_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_197_cast_fp16 = add(x = x_179_cast_fp16, y = var_4462_cast_fp16)[name = string("x_197_cast_fp16")];
+            int32 var_4469 = const()[name = string("op_4469"), val = int32(-1)];
+            fp16 const_118_promoted_to_fp16 = const()[name = string("const_118_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4475_cast_fp16 = mul(x = x_197_cast_fp16, y = const_118_promoted_to_fp16)[name = string("op_4475_cast_fp16")];
+            bool input_193_interleave_0 = const()[name = string("input_193_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_193_cast_fp16 = concat(axis = var_4469, interleave = input_193_interleave_0, values = (x_197_cast_fp16, var_4475_cast_fp16))[name = string("input_193_cast_fp16")];
+            tensor<int32, [1]> normed_185_axes_0 = const()[name = string("normed_185_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4467_to_fp16 = const()[name = string("op_4467_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_185_cast_fp16 = layer_norm(axes = normed_185_axes_0, epsilon = var_4467_to_fp16, x = input_193_cast_fp16)[name = string("normed_185_cast_fp16")];
+            tensor<int32, [2]> var_4480_split_sizes_0 = const()[name = string("op_4480_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4480_axis_0 = const()[name = string("op_4480_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4480_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4480_cast_fp16_1 = split(axis = var_4480_axis_0, split_sizes = var_4480_split_sizes_0, x = normed_185_cast_fp16)[name = string("op_4480_cast_fp16")];
+            tensor<fp16, [1536]> const_119_to_fp16 = const()[name = string("const_119_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214656000)))];
+            tensor<fp16, [1, 1, 1536]> var_4483_cast_fp16 = mul(x = var_4480_cast_fp16_0, y = const_119_to_fp16)[name = string("op_4483_cast_fp16")];
+            tensor<int32, [3]> var_4496 = const()[name = string("op_4496"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_195_axes_0 = const()[name = string("input_195_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4497 = transpose(perm = var_4496, x = var_4483_cast_fp16)[name = string("transpose_76")];
+            tensor<fp16, [1, 1536, 1, 1]> input_195 = expand_dims(axes = input_195_axes_0, x = var_4497)[name = string("input_195")];
+            string var_4510_pad_type_0 = const()[name = string("op_4510_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4510_strides_0 = const()[name = string("op_4510_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4510_pad_0 = const()[name = string("op_4510_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4510_dilations_0 = const()[name = string("op_4510_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4510_groups_0 = const()[name = string("op_4510_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_4510 = conv(dilations = var_4510_dilations_0, groups = var_4510_groups_0, pad = var_4510_pad_0, pad_type = var_4510_pad_type_0, strides = var_4510_strides_0, weight = layers_6_mlp_gate_proj_weight_palettized, x = input_195)[name = string("op_4510")];
+            string var_4512_mode_0 = const()[name = string("op_4512_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_4512 = gelu(mode = var_4512_mode_0, x = var_4510)[name = string("op_4512")];
+            string var_4523_pad_type_0 = const()[name = string("op_4523_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4523_strides_0 = const()[name = string("op_4523_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4523_pad_0 = const()[name = string("op_4523_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4523_dilations_0 = const()[name = string("op_4523_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4523_groups_0 = const()[name = string("op_4523_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_4523 = conv(dilations = var_4523_dilations_0, groups = var_4523_groups_0, pad = var_4523_pad_0, pad_type = var_4523_pad_type_0, strides = var_4523_strides_0, weight = layers_6_mlp_up_proj_weight_palettized, x = input_195)[name = string("op_4523")];
+            tensor<fp16, [1, 6144, 1, 1]> input_197 = mul(x = var_4512, y = var_4523)[name = string("input_197")];
+            string var_4535_pad_type_0 = const()[name = string("op_4535_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4535_strides_0 = const()[name = string("op_4535_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4535_pad_0 = const()[name = string("op_4535_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4535_dilations_0 = const()[name = string("op_4535_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4535_groups_0 = const()[name = string("op_4535_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_4535 = conv(dilations = var_4535_dilations_0, groups = var_4535_groups_0, pad = var_4535_pad_0, pad_type = var_4535_pad_type_0, strides = var_4535_strides_0, weight = layers_6_mlp_down_proj_weight_palettized, x = input_197)[name = string("op_4535")];
+            tensor<int32, [1]> var_4537_axes_0 = const()[name = string("op_4537_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4537 = squeeze(axes = var_4537_axes_0, x = var_4535)[name = string("op_4537")];
+            tensor<int32, [3]> var_4541 = const()[name = string("op_4541"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_4547 = const()[name = string("op_4547"), val = int32(-1)];
+            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_201 = transpose(perm = var_4541, x = var_4537)[name = string("transpose_75")];
+            tensor<fp16, [1, 1, 1536]> var_4553_cast_fp16 = mul(x = x_201, y = const_120_promoted_to_fp16)[name = string("op_4553_cast_fp16")];
+            bool input_199_interleave_0 = const()[name = string("input_199_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_199_cast_fp16 = concat(axis = var_4547, interleave = input_199_interleave_0, values = (x_201, var_4553_cast_fp16))[name = string("input_199_cast_fp16")];
+            tensor<int32, [1]> normed_189_axes_0 = const()[name = string("normed_189_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4545_to_fp16 = const()[name = string("op_4545_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_189_cast_fp16 = layer_norm(axes = normed_189_axes_0, epsilon = var_4545_to_fp16, x = input_199_cast_fp16)[name = string("normed_189_cast_fp16")];
+            tensor<int32, [2]> var_4558_split_sizes_0 = const()[name = string("op_4558_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4558_axis_0 = const()[name = string("op_4558_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4558_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4558_cast_fp16_1 = split(axis = var_4558_axis_0, split_sizes = var_4558_split_sizes_0, x = normed_189_cast_fp16)[name = string("op_4558_cast_fp16")];
+            tensor<fp16, [1536]> const_121_to_fp16 = const()[name = string("const_121_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214659136)))];
+            tensor<fp16, [1, 1, 1536]> var_4561_cast_fp16 = mul(x = var_4558_cast_fp16_0, y = const_121_to_fp16)[name = string("op_4561_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_97_cast_fp16 = add(x = x_197_cast_fp16, y = var_4561_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_4572 = linear(bias = linear_0_bias_0, weight = layers_6_per_layer_input_gate_weight_palettized, x = hidden_states_97_cast_fp16)[name = string("linear_12")];
+            string gated_13_mode_0 = const()[name = string("gated_13_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_13 = gelu(mode = gated_13_mode_0, x = var_4572)[name = string("gated_13")];
+            tensor<int32, [3]> var_4589_begin_0 = const()[name = string("op_4589_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
+            tensor<int32, [3]> var_4589_end_0 = const()[name = string("op_4589_end_0"), val = tensor<int32, [3]>([1, 1, 1792])];
+            tensor<bool, [3]> var_4589_end_mask_0 = const()[name = string("op_4589_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_4589_cast_fp16 = slice_by_index(begin = var_4589_begin_0, end = var_4589_end_0, end_mask = var_4589_end_mask_0, x = per_layer_combined)[name = string("op_4589_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_203_cast_fp16 = mul(x = gated_13, y = var_4589_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_6_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214662272))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214858944))))[name = string("layers_6_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_203_cast_fp16)[name = string("linear_13_cast_fp16")];
+            int32 var_4598 = const()[name = string("op_4598"), val = int32(-1)];
+            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4604_cast_fp16 = mul(x = linear_13_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4604_cast_fp16")];
+            bool input_205_interleave_0 = const()[name = string("input_205_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_205_cast_fp16 = concat(axis = var_4598, interleave = input_205_interleave_0, values = (linear_13_cast_fp16, var_4604_cast_fp16))[name = string("input_205_cast_fp16")];
+            tensor<int32, [1]> normed_193_axes_0 = const()[name = string("normed_193_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4596_to_fp16 = const()[name = string("op_4596_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_193_cast_fp16 = layer_norm(axes = normed_193_axes_0, epsilon = var_4596_to_fp16, x = input_205_cast_fp16)[name = string("normed_193_cast_fp16")];
+            tensor<int32, [2]> var_4609_split_sizes_0 = const()[name = string("op_4609_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4609_axis_0 = const()[name = string("op_4609_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4609_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4609_cast_fp16_1 = split(axis = var_4609_axis_0, split_sizes = var_4609_split_sizes_0, x = normed_193_cast_fp16)[name = string("op_4609_cast_fp16")];
+            tensor<fp16, [1536]> const_123_to_fp16 = const()[name = string("const_123_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214860544)))];
+            tensor<fp16, [1, 1, 1536]> var_4612_cast_fp16 = mul(x = var_4609_cast_fp16_0, y = const_123_to_fp16)[name = string("op_4612_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_101_cast_fp16 = add(x = hidden_states_97_cast_fp16, y = var_4612_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
+            tensor<fp16, [1]> layers_6_layer_scalar_to_fp16 = const()[name = string("layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fep-2])];
+            tensor<fp16, [1, 1, 1536]> x_209_cast_fp16 = mul(x = hidden_states_101_cast_fp16, y = layers_6_layer_scalar_to_fp16)[name = string("x_209_cast_fp16")];
+            int32 var_4620 = const()[name = string("op_4620"), val = int32(-1)];
+            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_4626_cast_fp16 = mul(x = x_209_cast_fp16, y = const_124_promoted_to_fp16)[name = string("op_4626_cast_fp16")];
+            bool input_207_interleave_0 = const()[name = string("input_207_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_207_cast_fp16 = concat(axis = var_4620, interleave = input_207_interleave_0, values = (x_209_cast_fp16, var_4626_cast_fp16))[name = string("input_207_cast_fp16")];
+            tensor<int32, [1]> normed_197_axes_0 = const()[name = string("normed_197_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4618_to_fp16 = const()[name = string("op_4618_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_197_cast_fp16 = layer_norm(axes = normed_197_axes_0, epsilon = var_4618_to_fp16, x = input_207_cast_fp16)[name = string("normed_197_cast_fp16")];
+            tensor<int32, [2]> var_4631_split_sizes_0 = const()[name = string("op_4631_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_4631_axis_0 = const()[name = string("op_4631_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_4631_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_4631_cast_fp16_1 = split(axis = var_4631_axis_0, split_sizes = var_4631_split_sizes_0, x = normed_197_cast_fp16)[name = string("op_4631_cast_fp16")];
+            tensor<fp16, [1536]> const_125_to_fp16 = const()[name = string("const_125_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214863680)))];
+            tensor<fp16, [1, 1, 1536]> var_4634_cast_fp16 = mul(x = var_4631_cast_fp16_0, y = const_125_to_fp16)[name = string("op_4634_cast_fp16")];
+            tensor<int32, [3]> var_4642 = const()[name = string("op_4642"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_4645_axes_0 = const()[name = string("op_4645_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_4643_cast_fp16 = transpose(perm = var_4642, x = var_4634_cast_fp16)[name = string("transpose_74")];
+            tensor<fp16, [1, 1536, 1, 1]> var_4645_cast_fp16 = expand_dims(axes = var_4645_axes_0, x = var_4643_cast_fp16)[name = string("op_4645_cast_fp16")];
+            string var_4661_pad_type_0 = const()[name = string("op_4661_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4661_strides_0 = const()[name = string("op_4661_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4661_pad_0 = const()[name = string("op_4661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4661_dilations_0 = const()[name = string("op_4661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4661_groups_0 = const()[name = string("op_4661_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_4661 = conv(dilations = var_4661_dilations_0, groups = var_4661_groups_0, pad = var_4661_pad_0, pad_type = var_4661_pad_type_0, strides = var_4661_strides_0, weight = layers_7_self_attn_q_proj_weight_palettized, x = var_4645_cast_fp16)[name = string("op_4661")];
+            tensor<int32, [4]> var_4666 = const()[name = string("op_4666"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_4667 = reshape(shape = var_4666, x = var_4661)[name = string("op_4667")];
+            tensor<int32, [4]> var_4672 = const()[name = string("op_4672"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_4682 = const()[name = string("op_4682"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_4673 = transpose(perm = var_4672, x = var_4667)[name = string("transpose_73")];
+            tensor<fp16, [1, 8, 256]> x_213 = reshape(shape = var_4682, x = var_4673)[name = string("x_213")];
+            int32 var_4688 = const()[name = string("op_4688"), val = int32(-1)];
+            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_4694_cast_fp16 = mul(x = x_213, y = const_126_promoted_to_fp16)[name = string("op_4694_cast_fp16")];
+            bool input_211_interleave_0 = const()[name = string("input_211_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_211_cast_fp16 = concat(axis = var_4688, interleave = input_211_interleave_0, values = (x_213, var_4694_cast_fp16))[name = string("input_211_cast_fp16")];
+            tensor<int32, [1]> normed_201_axes_0 = const()[name = string("normed_201_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_201_cast_fp16 = layer_norm(axes = normed_201_axes_0, epsilon = var_4686_to_fp16, x = input_211_cast_fp16)[name = string("normed_201_cast_fp16")];
+            tensor<int32, [2]> var_4699_split_sizes_0 = const()[name = string("op_4699_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_4699_axis_0 = const()[name = string("op_4699_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_4699_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_4699_cast_fp16_1 = split(axis = var_4699_axis_0, split_sizes = var_4699_split_sizes_0, x = normed_201_cast_fp16)[name = string("op_4699_cast_fp16")];
+            tensor<fp16, [256]> const_127_to_fp16 = const()[name = string("const_127_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214866816)))];
+            tensor<fp16, [1, 8, 256]> var_4702_cast_fp16 = mul(x = var_4699_cast_fp16_0, y = const_127_to_fp16)[name = string("op_4702_cast_fp16")];
+            tensor<int32, [4]> var_4708 = const()[name = string("op_4708"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_59 = reshape(shape = var_4708, x = var_4702_cast_fp16)[name = string("q_59")];
+            tensor<fp16, [1, 8, 1, 256]> var_4710 = mul(x = q_59, y = cos_1)[name = string("op_4710")];
+            tensor<int32, [2]> var_4711_split_sizes_0 = const()[name = string("op_4711_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_4711_axis_0 = const()[name = string("op_4711_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_4711_0, tensor<fp16, [1, 8, 1, 128]> var_4711_1 = split(axis = var_4711_axis_0, split_sizes = var_4711_split_sizes_0, x = q_59)[name = string("op_4711")];
+            fp16 const_128_promoted = const()[name = string("const_128_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_4713 = mul(x = var_4711_1, y = const_128_promoted)[name = string("op_4713")];
+            int32 var_4715 = const()[name = string("op_4715"), val = int32(-1)];
+            bool var_4716_interleave_0 = const()[name = string("op_4716_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_4716 = concat(axis = var_4715, interleave = var_4716_interleave_0, values = (var_4713, var_4711_0))[name = string("op_4716")];
+            tensor<fp16, [1, 8, 1, 256]> var_4717 = mul(x = var_4716, y = sin_1)[name = string("op_4717")];
+            tensor<fp16, [1, 8, 1, 256]> q_63 = add(x = var_4710, y = var_4717)[name = string("q_63")];
+            string var_4730_pad_type_0 = const()[name = string("op_4730_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4730_strides_0 = const()[name = string("op_4730_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4730_pad_0 = const()[name = string("op_4730_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4730_dilations_0 = const()[name = string("op_4730_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4730_groups_0 = const()[name = string("op_4730_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_4730 = conv(dilations = var_4730_dilations_0, groups = var_4730_groups_0, pad = var_4730_pad_0, pad_type = var_4730_pad_type_0, strides = var_4730_strides_0, weight = layers_7_self_attn_k_proj_weight_palettized, x = var_4645_cast_fp16)[name = string("op_4730")];
+            tensor<int32, [4]> var_4735 = const()[name = string("op_4735"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_4736 = reshape(shape = var_4735, x = var_4730)[name = string("op_4736")];
+            tensor<int32, [4]> var_4741 = const()[name = string("op_4741"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_4758_pad_type_0 = const()[name = string("op_4758_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_4758_strides_0 = const()[name = string("op_4758_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_4758_pad_0 = const()[name = string("op_4758_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_4758_dilations_0 = const()[name = string("op_4758_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_4758_groups_0 = const()[name = string("op_4758_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_4758 = conv(dilations = var_4758_dilations_0, groups = var_4758_groups_0, pad = var_4758_pad_0, pad_type = var_4758_pad_type_0, strides = var_4758_strides_0, weight = layers_7_self_attn_v_proj_weight_palettized, x = var_4645_cast_fp16)[name = string("op_4758")];
+            tensor<int32, [4]> var_4763 = const()[name = string("op_4763"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_4764 = reshape(shape = var_4763, x = var_4758)[name = string("op_4764")];
+            tensor<int32, [4]> var_4769 = const()[name = string("op_4769"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_4779 = const()[name = string("op_4779"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_4742 = transpose(perm = var_4741, x = var_4736)[name = string("transpose_72")];
+            tensor<fp16, [1, 1, 256]> x_217 = reshape(shape = var_4779, x = var_4742)[name = string("x_217")];
+            int32 var_4785 = const()[name = string("op_4785"), val = int32(-1)];
+            fp16 const_129_promoted_to_fp16 = const()[name = string("const_129_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_4791_cast_fp16 = mul(x = x_217, y = const_129_promoted_to_fp16)[name = string("op_4791_cast_fp16")];
+            bool input_213_interleave_0 = const()[name = string("input_213_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_213_cast_fp16 = concat(axis = var_4785, interleave = input_213_interleave_0, values = (x_217, var_4791_cast_fp16))[name = string("input_213_cast_fp16")];
+            tensor<int32, [1]> normed_205_axes_0 = const()[name = string("normed_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_4783_to_fp16 = const()[name = string("op_4783_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_205_cast_fp16 = layer_norm(axes = normed_205_axes_0, epsilon = var_4783_to_fp16, x = input_213_cast_fp16)[name = string("normed_205_cast_fp16")];
+            tensor<int32, [2]> var_4796_split_sizes_0 = const()[name = string("op_4796_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_4796_axis_0 = const()[name = string("op_4796_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_4796_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_4796_cast_fp16_1 = split(axis = var_4796_axis_0, split_sizes = var_4796_split_sizes_0, x = normed_205_cast_fp16)[name = string("op_4796_cast_fp16")];
+            tensor<fp16, [256]> const_130_to_fp16 = const()[name = string("const_130_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214867392)))];
+            tensor<fp16, [1, 1, 256]> var_4799_cast_fp16 = mul(x = var_4796_cast_fp16_0, y = const_130_to_fp16)[name = string("op_4799_cast_fp16")];
+            tensor<int32, [4]> var_4805 = const()[name = string("op_4805"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_61 = reshape(shape = var_4805, x = var_4799_cast_fp16)[name = string("q_61")];
+            fp16 var_4812_promoted_to_fp16 = const()[name = string("op_4812_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_4770 = transpose(perm = var_4769, x = var_4764)[name = string("transpose_71")];
+            tensor<fp16, [1, 1, 1, 256]> var_4813_cast_fp16 = pow(x = var_4770, y = var_4812_promoted_to_fp16)[name = string("op_4813_cast_fp16")];
+            tensor<int32, [1]> var_4818_axes_0 = const()[name = string("op_4818_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_4818_keep_dims_0 = const()[name = string("op_4818_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_4818_cast_fp16 = reduce_mean(axes = var_4818_axes_0, keep_dims = var_4818_keep_dims_0, x = var_4813_cast_fp16)[name = string("op_4818_cast_fp16")];
+            fp16 var_4820_to_fp16 = const()[name = string("op_4820_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_15_cast_fp16 = add(x = var_4818_cast_fp16, y = var_4820_to_fp16)[name = string("mean_sq_15_cast_fp16")];
+            fp16 var_4827_to_fp16 = const()[name = string("op_4827_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_4828_cast_fp16 = pow(x = mean_sq_15_cast_fp16, y = var_4827_to_fp16)[name = string("op_4828_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_4829_cast_fp16 = mul(x = var_4770, y = var_4828_cast_fp16)[name = string("op_4829_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_4835 = mul(x = q_61, y = cos_1)[name = string("op_4835")];
+            tensor<int32, [2]> var_4836_split_sizes_0 = const()[name = string("op_4836_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_4836_axis_0 = const()[name = string("op_4836_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_4836_0, tensor<fp16, [1, 1, 1, 128]> var_4836_1 = split(axis = var_4836_axis_0, split_sizes = var_4836_split_sizes_0, x = q_61)[name = string("op_4836")];
+            fp16 const_131_promoted = const()[name = string("const_131_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_4838 = mul(x = var_4836_1, y = const_131_promoted)[name = string("op_4838")];
+            int32 var_4840 = const()[name = string("op_4840"), val = int32(-1)];
+            bool var_4841_interleave_0 = const()[name = string("op_4841_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_4841 = concat(axis = var_4840, interleave = var_4841_interleave_0, values = (var_4838, var_4836_0))[name = string("op_4841")];
+            tensor<fp16, [1, 1, 1, 256]> var_4842 = mul(x = var_4841, y = sin_1)[name = string("op_4842")];
+            tensor<fp16, [1, 1, 1, 256]> input_215 = add(x = var_4835, y = var_4842)[name = string("input_215")];
+            tensor<int32, [4]> var_4847_begin_0 = const()[name = string("op_4847_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> var_4847_end_0 = const()[name = string("op_4847_end_0"), val = tensor<int32, [4]>([8, 1, 512, 512])];
+            tensor<bool, [4]> var_4847_end_mask_0 = const()[name = string("op_4847_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_4847_squeeze_mask_0 = const()[name = string("op_4847_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_4847_cast_fp16 = slice_by_index(begin = var_4847_begin_0, end = var_4847_end_0, end_mask = var_4847_end_mask_0, squeeze_mask = var_4847_squeeze_mask_0, x = coreml_update_state_37)[name = string("op_4847_cast_fp16")];
+            tensor<int32, [1]> K_c_15_axes_0 = const()[name = string("K_c_15_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_15_cast_fp16 = expand_dims(axes = K_c_15_axes_0, x = var_4847_cast_fp16)[name = string("K_c_15_cast_fp16")];
+            tensor<int32, [4]> var_4852_begin_0 = const()[name = string("op_4852_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> var_4852_end_0 = const()[name = string("op_4852_end_0"), val = tensor<int32, [4]>([20, 1, 512, 512])];
+            tensor<bool, [4]> var_4852_end_mask_0 = const()[name = string("op_4852_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_4852_squeeze_mask_0 = const()[name = string("op_4852_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_4852_cast_fp16 = slice_by_index(begin = var_4852_begin_0, end = var_4852_end_0, end_mask = var_4852_end_mask_0, squeeze_mask = var_4852_squeeze_mask_0, x = coreml_update_state_37)[name = string("op_4852_cast_fp16")];
+            tensor<int32, [1]> V_c_15_axes_0 = const()[name = string("V_c_15_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_15_cast_fp16 = expand_dims(axes = V_c_15_axes_0, x = var_4852_cast_fp16)[name = string("V_c_15_cast_fp16")];
+            tensor<int32, [8]> kp_13_pad_0 = const()[name = string("kp_13_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_13_mode_0 = const()[name = string("kp_13_mode_0"), val = string("constant")];
+            fp16 const_132_to_fp16 = const()[name = string("const_132_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_13_cast_fp16 = pad(constant_val = const_132_to_fp16, mode = kp_13_mode_0, pad = kp_13_pad_0, x = input_215)[name = string("kp_13_cast_fp16")];
+            tensor<int32, [8]> vp_13_pad_0 = const()[name = string("vp_13_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_13_mode_0 = const()[name = string("vp_13_mode_0"), val = string("constant")];
+            fp16 const_133_to_fp16 = const()[name = string("const_133_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_13_cast_fp16 = pad(constant_val = const_133_to_fp16, mode = vp_13_mode_0, pad = vp_13_pad_0, x = var_4829_cast_fp16)[name = string("vp_13_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4870_cast_fp16 = mul(x = K_c_15_cast_fp16, y = var_1005_cast_fp16)[name = string("op_4870_cast_fp16")];
+            tensor<int32, [4]> var_4871_reps_0 = const()[name = string("op_4871_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_4871_cast_fp16 = tile(reps = var_4871_reps_0, x = kp_13_cast_fp16)[name = string("op_4871_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4872_cast_fp16 = mul(x = var_4871_cast_fp16, y = update_mask)[name = string("op_4872_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_15_cast_fp16 = add(x = var_4870_cast_fp16, y = var_4872_cast_fp16)[name = string("K_n_15_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4878_cast_fp16 = mul(x = V_c_15_cast_fp16, y = var_1005_cast_fp16)[name = string("op_4878_cast_fp16")];
+            tensor<int32, [4]> var_4879_reps_0 = const()[name = string("op_4879_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_4879_cast_fp16 = tile(reps = var_4879_reps_0, x = vp_13_cast_fp16)[name = string("op_4879_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_4880_cast_fp16 = mul(x = var_4879_cast_fp16, y = update_mask)[name = string("op_4880_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_15_cast_fp16 = add(x = var_4878_cast_fp16, y = var_4880_cast_fp16)[name = string("V_n_15_cast_fp16")];
+            tensor<int32, [1]> var_4884_axes_0 = const()[name = string("op_4884_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_4884_cast_fp16 = squeeze(axes = var_4884_axes_0, x = K_n_15_cast_fp16)[name = string("op_4884_cast_fp16")];
+            tensor<int32, [4]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_15_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_15_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_15_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_56, begin_mask = kv_cache_0_internal_tensor_assign_15_begin_mask_0, end = concat_57, end_mask = kv_cache_0_internal_tensor_assign_15_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_15_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_15_stride_0, update = var_4884_cast_fp16, x = coreml_update_state_37)[name = string("kv_cache_0_internal_tensor_assign_15_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_15_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_38_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_38 = read_state(input = kv_cache_0)[name = string("coreml_update_state_38")];
+            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_4891_cast_fp16 = squeeze(axes = var_4891_axes_0, x = V_n_15_cast_fp16)[name = string("op_4891_cast_fp16")];
+            tensor<int32, [4]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_16_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_16_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_16_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_58, begin_mask = kv_cache_0_internal_tensor_assign_16_begin_mask_0, end = concat_59, end_mask = kv_cache_0_internal_tensor_assign_16_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_16_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_16_stride_0, update = var_4891_cast_fp16, x = coreml_update_state_38)[name = string("kv_cache_0_internal_tensor_assign_16_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_16_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_39_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_39 = read_state(input = kv_cache_0)[name = string("coreml_update_state_39")];
+            tensor<int32, [4]> var_4901_begin_0 = const()[name = string("op_4901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4901_end_0 = const()[name = string("op_4901_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_4901_end_mask_0 = const()[name = string("op_4901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_4901_cast_fp16 = slice_by_index(begin = var_4901_begin_0, end = var_4901_end_0, end_mask = var_4901_end_mask_0, x = K_n_15_cast_fp16)[name = string("op_4901_cast_fp16")];
+            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = var_4901_cast_fp16)[name = string("transpose_70")];
+            tensor<fp16, [8, 1, 512, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
+            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_28_cast_fp16 = reshape(shape = concat_60, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
+            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_69")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_29_cast_fp16 = reshape(shape = concat_61, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
+            tensor<int32, [4]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_4910_begin_0 = const()[name = string("op_4910_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4910_end_0 = const()[name = string("op_4910_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_4910_end_mask_0 = const()[name = string("op_4910_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_4910_cast_fp16 = slice_by_index(begin = var_4910_begin_0, end = var_4910_end_0, end_mask = var_4910_end_mask_0, x = V_n_15_cast_fp16)[name = string("op_4910_cast_fp16")];
+            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = var_4910_cast_fp16)[name = string("transpose_68")];
+            tensor<fp16, [8, 1, 512, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
+            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_30_cast_fp16 = reshape(shape = concat_62, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
+            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_67")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_31_cast_fp16 = reshape(shape = concat_63, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
+            tensor<int32, [4]> Ve_15_perm_0 = const()[name = string("Ve_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_4928_transpose_x_0 = const()[name = string("op_4928_transpose_x_0"), val = bool(false)];
+            bool var_4928_transpose_y_0 = const()[name = string("op_4928_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_66")];
+            tensor<fp16, [1, 8, 1, 512]> var_4928_cast_fp16 = matmul(transpose_x = var_4928_transpose_x_0, transpose_y = var_4928_transpose_y_0, x = q_63, y = transpose_55_cast_fp16)[name = string("op_4928_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_4935_cast_fp16 = add(x = var_4928_cast_fp16, y = causal_mask)[name = string("op_4935_cast_fp16")];
+            int32 var_4936 = const()[name = string("op_4936"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_4938_cast_fp16 = softmax(axis = var_4936, x = var_4935_cast_fp16)[name = string("op_4938_cast_fp16")];
+            bool var_4954_transpose_x_0 = const()[name = string("op_4954_transpose_x_0"), val = bool(false)];
+            bool var_4954_transpose_y_0 = const()[name = string("op_4954_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_15_cast_fp16 = transpose(perm = Ve_15_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_65")];
+            tensor<fp16, [1, 8, 1, 256]> var_4954_cast_fp16 = matmul(transpose_x = var_4954_transpose_x_0, transpose_y = var_4954_transpose_y_0, x = var_4938_cast_fp16, y = Ve_15_cast_fp16)[name = string("op_4954_cast_fp16")];
+            tensor<int32, [4]> var_4964 = const()[name = string("op_4964"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4971 = const()[name = string("op_4971"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_4965 = transpose(perm = var_4964, x = var_4954_cast_fp16)[name = string("transpose_64")];
+            tensor<fp16, [1, 1, 2048]> var_4972 = reshape(shape = var_4971, x = var_4965)[name = string("op_4972")];
+            tensor<int32, [3]> var_4976 = const()[name = string("op_4976"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_7_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214867968))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216440896))))[name = string("squeeze_7_palettized")];
+            string var_4992_pad_type_0 = const()[name = string("op_4992_pad_type_0"), val = string("valid")];
+            int32 var_4992_groups_0 = const()[name = string("op_4992_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_4992_strides_0 = const()[name = string("op_4992_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_4992_pad_0 = const()[name = string("op_4992_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_4992_dilations_0 = const()[name = string("op_4992_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_4977 = transpose(perm = var_4976, x = var_4972)[name = string("transpose_63")];
+            tensor<fp16, [1, 1536, 1]> var_4992 = conv(dilations = var_4992_dilations_0, groups = var_4992_groups_0, pad = var_4992_pad_0, pad_type = var_4992_pad_type_0, strides = var_4992_strides_0, weight = squeeze_7_palettized, x = var_4977)[name = string("op_4992")];
+            tensor<int32, [3]> var_4996 = const()[name = string("op_4996"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_5002 = const()[name = string("op_5002"), val = int32(-1)];
+            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_223 = transpose(perm = var_4996, x = var_4992)[name = string("transpose_62")];
+            tensor<fp16, [1, 1, 1536]> var_5008_cast_fp16 = mul(x = x_223, y = const_134_promoted_to_fp16)[name = string("op_5008_cast_fp16")];
+            bool input_221_interleave_0 = const()[name = string("input_221_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_221_cast_fp16 = concat(axis = var_5002, interleave = input_221_interleave_0, values = (x_223, var_5008_cast_fp16))[name = string("input_221_cast_fp16")];
+            tensor<int32, [1]> normed_209_axes_0 = const()[name = string("normed_209_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5000_to_fp16 = const()[name = string("op_5000_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_209_cast_fp16 = layer_norm(axes = normed_209_axes_0, epsilon = var_5000_to_fp16, x = input_221_cast_fp16)[name = string("normed_209_cast_fp16")];
+            tensor<int32, [2]> var_5013_split_sizes_0 = const()[name = string("op_5013_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5013_axis_0 = const()[name = string("op_5013_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5013_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5013_cast_fp16_1 = split(axis = var_5013_axis_0, split_sizes = var_5013_split_sizes_0, x = normed_209_cast_fp16)[name = string("op_5013_cast_fp16")];
+            tensor<fp16, [1536]> const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216442496)))];
+            tensor<fp16, [1, 1, 1536]> var_5016_cast_fp16 = mul(x = var_5013_cast_fp16_0, y = const_135_to_fp16)[name = string("op_5016_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_227_cast_fp16 = add(x = x_209_cast_fp16, y = var_5016_cast_fp16)[name = string("x_227_cast_fp16")];
+            int32 var_5023 = const()[name = string("op_5023"), val = int32(-1)];
+            fp16 const_136_promoted_to_fp16 = const()[name = string("const_136_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_5029_cast_fp16 = mul(x = x_227_cast_fp16, y = const_136_promoted_to_fp16)[name = string("op_5029_cast_fp16")];
+            bool input_223_interleave_0 = const()[name = string("input_223_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_223_cast_fp16 = concat(axis = var_5023, interleave = input_223_interleave_0, values = (x_227_cast_fp16, var_5029_cast_fp16))[name = string("input_223_cast_fp16")];
+            tensor<int32, [1]> normed_213_axes_0 = const()[name = string("normed_213_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5021_to_fp16 = const()[name = string("op_5021_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_213_cast_fp16 = layer_norm(axes = normed_213_axes_0, epsilon = var_5021_to_fp16, x = input_223_cast_fp16)[name = string("normed_213_cast_fp16")];
+            tensor<int32, [2]> var_5034_split_sizes_0 = const()[name = string("op_5034_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5034_axis_0 = const()[name = string("op_5034_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5034_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5034_cast_fp16_1 = split(axis = var_5034_axis_0, split_sizes = var_5034_split_sizes_0, x = normed_213_cast_fp16)[name = string("op_5034_cast_fp16")];
+            tensor<fp16, [1536]> const_137_to_fp16 = const()[name = string("const_137_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216445632)))];
+            tensor<fp16, [1, 1, 1536]> var_5037_cast_fp16 = mul(x = var_5034_cast_fp16_0, y = const_137_to_fp16)[name = string("op_5037_cast_fp16")];
+            tensor<int32, [3]> var_5050 = const()[name = string("op_5050"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_225_axes_0 = const()[name = string("input_225_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_5051 = transpose(perm = var_5050, x = var_5037_cast_fp16)[name = string("transpose_61")];
+            tensor<fp16, [1, 1536, 1, 1]> input_225 = expand_dims(axes = input_225_axes_0, x = var_5051)[name = string("input_225")];
+            string var_5064_pad_type_0 = const()[name = string("op_5064_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5064_strides_0 = const()[name = string("op_5064_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5064_pad_0 = const()[name = string("op_5064_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5064_dilations_0 = const()[name = string("op_5064_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5064_groups_0 = const()[name = string("op_5064_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_5064 = conv(dilations = var_5064_dilations_0, groups = var_5064_groups_0, pad = var_5064_pad_0, pad_type = var_5064_pad_type_0, strides = var_5064_strides_0, weight = layers_7_mlp_gate_proj_weight_palettized, x = input_225)[name = string("op_5064")];
+            string var_5066_mode_0 = const()[name = string("op_5066_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_5066 = gelu(mode = var_5066_mode_0, x = var_5064)[name = string("op_5066")];
+            string var_5077_pad_type_0 = const()[name = string("op_5077_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5077_strides_0 = const()[name = string("op_5077_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5077_pad_0 = const()[name = string("op_5077_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5077_dilations_0 = const()[name = string("op_5077_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5077_groups_0 = const()[name = string("op_5077_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_5077 = conv(dilations = var_5077_dilations_0, groups = var_5077_groups_0, pad = var_5077_pad_0, pad_type = var_5077_pad_type_0, strides = var_5077_strides_0, weight = layers_7_mlp_up_proj_weight_palettized, x = input_225)[name = string("op_5077")];
+            tensor<fp16, [1, 6144, 1, 1]> input_227 = mul(x = var_5066, y = var_5077)[name = string("input_227")];
+            string var_5089_pad_type_0 = const()[name = string("op_5089_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5089_strides_0 = const()[name = string("op_5089_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5089_pad_0 = const()[name = string("op_5089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5089_dilations_0 = const()[name = string("op_5089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5089_groups_0 = const()[name = string("op_5089_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_5089 = conv(dilations = var_5089_dilations_0, groups = var_5089_groups_0, pad = var_5089_pad_0, pad_type = var_5089_pad_type_0, strides = var_5089_strides_0, weight = layers_7_mlp_down_proj_weight_palettized, x = input_227)[name = string("op_5089")];
+            tensor<int32, [1]> var_5091_axes_0 = const()[name = string("op_5091_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_5091 = squeeze(axes = var_5091_axes_0, x = var_5089)[name = string("op_5091")];
+            tensor<int32, [3]> var_5095 = const()[name = string("op_5095"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_5101 = const()[name = string("op_5101"), val = int32(-1)];
+            fp16 const_138_promoted_to_fp16 = const()[name = string("const_138_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_231 = transpose(perm = var_5095, x = var_5091)[name = string("transpose_60")];
+            tensor<fp16, [1, 1, 1536]> var_5107_cast_fp16 = mul(x = x_231, y = const_138_promoted_to_fp16)[name = string("op_5107_cast_fp16")];
+            bool input_229_interleave_0 = const()[name = string("input_229_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_229_cast_fp16 = concat(axis = var_5101, interleave = input_229_interleave_0, values = (x_231, var_5107_cast_fp16))[name = string("input_229_cast_fp16")];
+            tensor<int32, [1]> normed_217_axes_0 = const()[name = string("normed_217_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5099_to_fp16 = const()[name = string("op_5099_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_217_cast_fp16 = layer_norm(axes = normed_217_axes_0, epsilon = var_5099_to_fp16, x = input_229_cast_fp16)[name = string("normed_217_cast_fp16")];
+            tensor<int32, [2]> var_5112_split_sizes_0 = const()[name = string("op_5112_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5112_axis_0 = const()[name = string("op_5112_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5112_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5112_cast_fp16_1 = split(axis = var_5112_axis_0, split_sizes = var_5112_split_sizes_0, x = normed_217_cast_fp16)[name = string("op_5112_cast_fp16")];
+            tensor<fp16, [1536]> const_139_to_fp16 = const()[name = string("const_139_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216448768)))];
+            tensor<fp16, [1, 1, 1536]> var_5115_cast_fp16 = mul(x = var_5112_cast_fp16_0, y = const_139_to_fp16)[name = string("op_5115_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_111_cast_fp16 = add(x = x_227_cast_fp16, y = var_5115_cast_fp16)[name = string("hidden_states_111_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_5126 = linear(bias = linear_0_bias_0, weight = layers_7_per_layer_input_gate_weight_palettized, x = hidden_states_111_cast_fp16)[name = string("linear_14")];
+            string gated_15_mode_0 = const()[name = string("gated_15_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_15 = gelu(mode = gated_15_mode_0, x = var_5126)[name = string("gated_15")];
+            tensor<int32, [3]> var_5143_begin_0 = const()[name = string("op_5143_begin_0"), val = tensor<int32, [3]>([0, 0, 1792])];
+            tensor<int32, [3]> var_5143_end_0 = const()[name = string("op_5143_end_0"), val = tensor<int32, [3]>([1, 1, 2048])];
+            tensor<bool, [3]> var_5143_end_mask_0 = const()[name = string("op_5143_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_5143_cast_fp16 = slice_by_index(begin = var_5143_begin_0, end = var_5143_end_0, end_mask = var_5143_end_mask_0, x = per_layer_combined)[name = string("op_5143_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_233_cast_fp16 = mul(x = gated_15, y = var_5143_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_7_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216451904))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216648576))))[name = string("layers_7_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_233_cast_fp16)[name = string("linear_15_cast_fp16")];
+            int32 var_5152 = const()[name = string("op_5152"), val = int32(-1)];
+            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_5158_cast_fp16 = mul(x = linear_15_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_5158_cast_fp16")];
+            bool input_235_interleave_0 = const()[name = string("input_235_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_235_cast_fp16 = concat(axis = var_5152, interleave = input_235_interleave_0, values = (linear_15_cast_fp16, var_5158_cast_fp16))[name = string("input_235_cast_fp16")];
+            tensor<int32, [1]> normed_221_axes_0 = const()[name = string("normed_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5150_to_fp16 = const()[name = string("op_5150_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_221_cast_fp16 = layer_norm(axes = normed_221_axes_0, epsilon = var_5150_to_fp16, x = input_235_cast_fp16)[name = string("normed_221_cast_fp16")];
+            tensor<int32, [2]> var_5163_split_sizes_0 = const()[name = string("op_5163_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5163_axis_0 = const()[name = string("op_5163_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5163_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5163_cast_fp16_1 = split(axis = var_5163_axis_0, split_sizes = var_5163_split_sizes_0, x = normed_221_cast_fp16)[name = string("op_5163_cast_fp16")];
+            tensor<fp16, [1536]> const_141_to_fp16 = const()[name = string("const_141_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216650176)))];
+            tensor<fp16, [1, 1, 1536]> var_5166_cast_fp16 = mul(x = var_5163_cast_fp16_0, y = const_141_to_fp16)[name = string("op_5166_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_115_cast_fp16 = add(x = hidden_states_111_cast_fp16, y = var_5166_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
+            tensor<fp16, [1]> layers_7_layer_scalar_to_fp16 = const()[name = string("layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.38p-1])];
+            tensor<fp16, [1, 1, 1536]> x_239_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = layers_7_layer_scalar_to_fp16)[name = string("x_239_cast_fp16")];
+            int32 var_5174 = const()[name = string("op_5174"), val = int32(-1)];
+            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_5180_cast_fp16 = mul(x = x_239_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_5180_cast_fp16")];
+            bool input_237_interleave_0 = const()[name = string("input_237_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_237_cast_fp16 = concat(axis = var_5174, interleave = input_237_interleave_0, values = (x_239_cast_fp16, var_5180_cast_fp16))[name = string("input_237_cast_fp16")];
+            tensor<int32, [1]> normed_225_axes_0 = const()[name = string("normed_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5172_to_fp16 = const()[name = string("op_5172_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_225_cast_fp16 = layer_norm(axes = normed_225_axes_0, epsilon = var_5172_to_fp16, x = input_237_cast_fp16)[name = string("normed_225_cast_fp16")];
+            tensor<int32, [2]> var_5185_split_sizes_0 = const()[name = string("op_5185_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5185_axis_0 = const()[name = string("op_5185_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5185_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5185_cast_fp16_1 = split(axis = var_5185_axis_0, split_sizes = var_5185_split_sizes_0, x = normed_225_cast_fp16)[name = string("op_5185_cast_fp16")];
+            tensor<fp16, [1536]> const_143_to_fp16 = const()[name = string("const_143_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216653312)))];
+            tensor<fp16, [1, 1, 1536]> var_5188_cast_fp16 = mul(x = var_5185_cast_fp16_0, y = const_143_to_fp16)[name = string("op_5188_cast_fp16")];
+            tensor<int32, [3]> var_5196 = const()[name = string("op_5196"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_5199_axes_0 = const()[name = string("op_5199_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_5197_cast_fp16 = transpose(perm = var_5196, x = var_5188_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [1, 1536, 1, 1]> var_5199_cast_fp16 = expand_dims(axes = var_5199_axes_0, x = var_5197_cast_fp16)[name = string("op_5199_cast_fp16")];
+            string var_5215_pad_type_0 = const()[name = string("op_5215_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5215_strides_0 = const()[name = string("op_5215_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5215_pad_0 = const()[name = string("op_5215_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5215_dilations_0 = const()[name = string("op_5215_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5215_groups_0 = const()[name = string("op_5215_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_5215 = conv(dilations = var_5215_dilations_0, groups = var_5215_groups_0, pad = var_5215_pad_0, pad_type = var_5215_pad_type_0, strides = var_5215_strides_0, weight = layers_8_self_attn_q_proj_weight_palettized, x = var_5199_cast_fp16)[name = string("op_5215")];
+            tensor<int32, [4]> var_5220 = const()[name = string("op_5220"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_5221 = reshape(shape = var_5220, x = var_5215)[name = string("op_5221")];
+            tensor<int32, [4]> var_5226 = const()[name = string("op_5226"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_5236 = const()[name = string("op_5236"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_5227 = transpose(perm = var_5226, x = var_5221)[name = string("transpose_58")];
+            tensor<fp16, [1, 8, 256]> x_243 = reshape(shape = var_5236, x = var_5227)[name = string("x_243")];
+            int32 var_5242 = const()[name = string("op_5242"), val = int32(-1)];
+            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_5248_cast_fp16 = mul(x = x_243, y = const_144_promoted_to_fp16)[name = string("op_5248_cast_fp16")];
+            bool input_241_interleave_0 = const()[name = string("input_241_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_241_cast_fp16 = concat(axis = var_5242, interleave = input_241_interleave_0, values = (x_243, var_5248_cast_fp16))[name = string("input_241_cast_fp16")];
+            tensor<int32, [1]> normed_229_axes_0 = const()[name = string("normed_229_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5240_to_fp16 = const()[name = string("op_5240_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_229_cast_fp16 = layer_norm(axes = normed_229_axes_0, epsilon = var_5240_to_fp16, x = input_241_cast_fp16)[name = string("normed_229_cast_fp16")];
+            tensor<int32, [2]> var_5253_split_sizes_0 = const()[name = string("op_5253_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_5253_axis_0 = const()[name = string("op_5253_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_5253_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_5253_cast_fp16_1 = split(axis = var_5253_axis_0, split_sizes = var_5253_split_sizes_0, x = normed_229_cast_fp16)[name = string("op_5253_cast_fp16")];
+            tensor<fp16, [256]> const_145_to_fp16 = const()[name = string("const_145_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216656448)))];
+            tensor<fp16, [1, 8, 256]> var_5256_cast_fp16 = mul(x = var_5253_cast_fp16_0, y = const_145_to_fp16)[name = string("op_5256_cast_fp16")];
+            tensor<int32, [4]> var_5262 = const()[name = string("op_5262"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_67 = reshape(shape = var_5262, x = var_5256_cast_fp16)[name = string("q_67")];
+            tensor<fp16, [1, 8, 1, 256]> var_5264 = mul(x = q_67, y = cos_1)[name = string("op_5264")];
+            tensor<int32, [2]> var_5265_split_sizes_0 = const()[name = string("op_5265_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_5265_axis_0 = const()[name = string("op_5265_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_5265_0, tensor<fp16, [1, 8, 1, 128]> var_5265_1 = split(axis = var_5265_axis_0, split_sizes = var_5265_split_sizes_0, x = q_67)[name = string("op_5265")];
+            fp16 const_146_promoted = const()[name = string("const_146_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_5267 = mul(x = var_5265_1, y = const_146_promoted)[name = string("op_5267")];
+            int32 var_5269 = const()[name = string("op_5269"), val = int32(-1)];
+            bool var_5270_interleave_0 = const()[name = string("op_5270_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_5270 = concat(axis = var_5269, interleave = var_5270_interleave_0, values = (var_5267, var_5265_0))[name = string("op_5270")];
+            tensor<fp16, [1, 8, 1, 256]> var_5271 = mul(x = var_5270, y = sin_1)[name = string("op_5271")];
+            tensor<fp16, [1, 8, 1, 256]> q_71 = add(x = var_5264, y = var_5271)[name = string("q_71")];
+            string var_5284_pad_type_0 = const()[name = string("op_5284_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5284_strides_0 = const()[name = string("op_5284_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5284_pad_0 = const()[name = string("op_5284_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5284_dilations_0 = const()[name = string("op_5284_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5284_groups_0 = const()[name = string("op_5284_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_5284 = conv(dilations = var_5284_dilations_0, groups = var_5284_groups_0, pad = var_5284_pad_0, pad_type = var_5284_pad_type_0, strides = var_5284_strides_0, weight = layers_8_self_attn_k_proj_weight_palettized, x = var_5199_cast_fp16)[name = string("op_5284")];
+            tensor<int32, [4]> var_5289 = const()[name = string("op_5289"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_5290 = reshape(shape = var_5289, x = var_5284)[name = string("op_5290")];
+            tensor<int32, [4]> var_5295 = const()[name = string("op_5295"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_5312_pad_type_0 = const()[name = string("op_5312_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5312_strides_0 = const()[name = string("op_5312_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5312_pad_0 = const()[name = string("op_5312_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5312_dilations_0 = const()[name = string("op_5312_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5312_groups_0 = const()[name = string("op_5312_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_5312 = conv(dilations = var_5312_dilations_0, groups = var_5312_groups_0, pad = var_5312_pad_0, pad_type = var_5312_pad_type_0, strides = var_5312_strides_0, weight = layers_8_self_attn_v_proj_weight_palettized, x = var_5199_cast_fp16)[name = string("op_5312")];
+            tensor<int32, [4]> var_5317 = const()[name = string("op_5317"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_5318 = reshape(shape = var_5317, x = var_5312)[name = string("op_5318")];
+            tensor<int32, [4]> var_5323 = const()[name = string("op_5323"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_5333 = const()[name = string("op_5333"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_5296 = transpose(perm = var_5295, x = var_5290)[name = string("transpose_57")];
+            tensor<fp16, [1, 1, 256]> x_247 = reshape(shape = var_5333, x = var_5296)[name = string("x_247")];
+            int32 var_5339 = const()[name = string("op_5339"), val = int32(-1)];
+            fp16 const_147_promoted_to_fp16 = const()[name = string("const_147_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_5345_cast_fp16 = mul(x = x_247, y = const_147_promoted_to_fp16)[name = string("op_5345_cast_fp16")];
+            bool input_243_interleave_0 = const()[name = string("input_243_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_243_cast_fp16 = concat(axis = var_5339, interleave = input_243_interleave_0, values = (x_247, var_5345_cast_fp16))[name = string("input_243_cast_fp16")];
+            tensor<int32, [1]> normed_233_axes_0 = const()[name = string("normed_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5337_to_fp16 = const()[name = string("op_5337_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_233_cast_fp16 = layer_norm(axes = normed_233_axes_0, epsilon = var_5337_to_fp16, x = input_243_cast_fp16)[name = string("normed_233_cast_fp16")];
+            tensor<int32, [2]> var_5350_split_sizes_0 = const()[name = string("op_5350_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_5350_axis_0 = const()[name = string("op_5350_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_5350_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_5350_cast_fp16_1 = split(axis = var_5350_axis_0, split_sizes = var_5350_split_sizes_0, x = normed_233_cast_fp16)[name = string("op_5350_cast_fp16")];
+            tensor<fp16, [256]> const_148_to_fp16 = const()[name = string("const_148_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216657024)))];
+            tensor<fp16, [1, 1, 256]> var_5353_cast_fp16 = mul(x = var_5350_cast_fp16_0, y = const_148_to_fp16)[name = string("op_5353_cast_fp16")];
+            tensor<int32, [4]> var_5359 = const()[name = string("op_5359"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_69 = reshape(shape = var_5359, x = var_5353_cast_fp16)[name = string("q_69")];
+            fp16 var_5366_promoted_to_fp16 = const()[name = string("op_5366_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_5324 = transpose(perm = var_5323, x = var_5318)[name = string("transpose_56")];
+            tensor<fp16, [1, 1, 1, 256]> var_5367_cast_fp16 = pow(x = var_5324, y = var_5366_promoted_to_fp16)[name = string("op_5367_cast_fp16")];
+            tensor<int32, [1]> var_5372_axes_0 = const()[name = string("op_5372_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_5372_keep_dims_0 = const()[name = string("op_5372_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_5372_cast_fp16 = reduce_mean(axes = var_5372_axes_0, keep_dims = var_5372_keep_dims_0, x = var_5367_cast_fp16)[name = string("op_5372_cast_fp16")];
+            fp16 var_5374_to_fp16 = const()[name = string("op_5374_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_17_cast_fp16 = add(x = var_5372_cast_fp16, y = var_5374_to_fp16)[name = string("mean_sq_17_cast_fp16")];
+            fp16 var_5381_to_fp16 = const()[name = string("op_5381_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_5382_cast_fp16 = pow(x = mean_sq_17_cast_fp16, y = var_5381_to_fp16)[name = string("op_5382_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_5383_cast_fp16 = mul(x = var_5324, y = var_5382_cast_fp16)[name = string("op_5383_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_5389 = mul(x = q_69, y = cos_1)[name = string("op_5389")];
+            tensor<int32, [2]> var_5390_split_sizes_0 = const()[name = string("op_5390_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_5390_axis_0 = const()[name = string("op_5390_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_5390_0, tensor<fp16, [1, 1, 1, 128]> var_5390_1 = split(axis = var_5390_axis_0, split_sizes = var_5390_split_sizes_0, x = q_69)[name = string("op_5390")];
+            fp16 const_149_promoted = const()[name = string("const_149_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_5392 = mul(x = var_5390_1, y = const_149_promoted)[name = string("op_5392")];
+            int32 var_5394 = const()[name = string("op_5394"), val = int32(-1)];
+            bool var_5395_interleave_0 = const()[name = string("op_5395_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_5395 = concat(axis = var_5394, interleave = var_5395_interleave_0, values = (var_5392, var_5390_0))[name = string("op_5395")];
+            tensor<fp16, [1, 1, 1, 256]> var_5396 = mul(x = var_5395, y = sin_1)[name = string("op_5396")];
+            tensor<fp16, [1, 1, 1, 256]> input_245 = add(x = var_5389, y = var_5396)[name = string("input_245")];
+            tensor<int32, [4]> var_5401_begin_0 = const()[name = string("op_5401_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> var_5401_end_0 = const()[name = string("op_5401_end_0"), val = tensor<int32, [4]>([9, 1, 512, 512])];
+            tensor<bool, [4]> var_5401_end_mask_0 = const()[name = string("op_5401_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_5401_squeeze_mask_0 = const()[name = string("op_5401_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_5401_cast_fp16 = slice_by_index(begin = var_5401_begin_0, end = var_5401_end_0, end_mask = var_5401_end_mask_0, squeeze_mask = var_5401_squeeze_mask_0, x = coreml_update_state_39)[name = string("op_5401_cast_fp16")];
+            tensor<int32, [1]> K_c_17_axes_0 = const()[name = string("K_c_17_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_17_cast_fp16 = expand_dims(axes = K_c_17_axes_0, x = var_5401_cast_fp16)[name = string("K_c_17_cast_fp16")];
+            tensor<int32, [4]> var_5406_begin_0 = const()[name = string("op_5406_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> var_5406_end_0 = const()[name = string("op_5406_end_0"), val = tensor<int32, [4]>([21, 1, 512, 512])];
+            tensor<bool, [4]> var_5406_end_mask_0 = const()[name = string("op_5406_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_5406_squeeze_mask_0 = const()[name = string("op_5406_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_5406_cast_fp16 = slice_by_index(begin = var_5406_begin_0, end = var_5406_end_0, end_mask = var_5406_end_mask_0, squeeze_mask = var_5406_squeeze_mask_0, x = coreml_update_state_39)[name = string("op_5406_cast_fp16")];
+            tensor<int32, [1]> V_c_17_axes_0 = const()[name = string("V_c_17_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_17_cast_fp16 = expand_dims(axes = V_c_17_axes_0, x = var_5406_cast_fp16)[name = string("V_c_17_cast_fp16")];
+            tensor<int32, [8]> kp_15_pad_0 = const()[name = string("kp_15_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_15_mode_0 = const()[name = string("kp_15_mode_0"), val = string("constant")];
+            fp16 const_150_to_fp16 = const()[name = string("const_150_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_15_cast_fp16 = pad(constant_val = const_150_to_fp16, mode = kp_15_mode_0, pad = kp_15_pad_0, x = input_245)[name = string("kp_15_cast_fp16")];
+            tensor<int32, [8]> vp_15_pad_0 = const()[name = string("vp_15_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_15_mode_0 = const()[name = string("vp_15_mode_0"), val = string("constant")];
+            fp16 const_151_to_fp16 = const()[name = string("const_151_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_15_cast_fp16 = pad(constant_val = const_151_to_fp16, mode = vp_15_mode_0, pad = vp_15_pad_0, x = var_5383_cast_fp16)[name = string("vp_15_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_5424_cast_fp16 = mul(x = K_c_17_cast_fp16, y = var_1005_cast_fp16)[name = string("op_5424_cast_fp16")];
+            tensor<int32, [4]> var_5425_reps_0 = const()[name = string("op_5425_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_5425_cast_fp16 = tile(reps = var_5425_reps_0, x = kp_15_cast_fp16)[name = string("op_5425_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_5426_cast_fp16 = mul(x = var_5425_cast_fp16, y = update_mask)[name = string("op_5426_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_17_cast_fp16 = add(x = var_5424_cast_fp16, y = var_5426_cast_fp16)[name = string("K_n_17_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_5432_cast_fp16 = mul(x = V_c_17_cast_fp16, y = var_1005_cast_fp16)[name = string("op_5432_cast_fp16")];
+            tensor<int32, [4]> var_5433_reps_0 = const()[name = string("op_5433_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_5433_cast_fp16 = tile(reps = var_5433_reps_0, x = vp_15_cast_fp16)[name = string("op_5433_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_5434_cast_fp16 = mul(x = var_5433_cast_fp16, y = update_mask)[name = string("op_5434_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_17_cast_fp16 = add(x = var_5432_cast_fp16, y = var_5434_cast_fp16)[name = string("V_n_17_cast_fp16")];
+            tensor<int32, [1]> var_5438_axes_0 = const()[name = string("op_5438_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_5438_cast_fp16 = squeeze(axes = var_5438_axes_0, x = K_n_17_cast_fp16)[name = string("op_5438_cast_fp16")];
+            tensor<int32, [4]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_17_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_17_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_17_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_64, begin_mask = kv_cache_0_internal_tensor_assign_17_begin_mask_0, end = concat_65, end_mask = kv_cache_0_internal_tensor_assign_17_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_17_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_17_stride_0, update = var_5438_cast_fp16, x = coreml_update_state_39)[name = string("kv_cache_0_internal_tensor_assign_17_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_17_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_40_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_40 = read_state(input = kv_cache_0)[name = string("coreml_update_state_40")];
+            tensor<int32, [1]> var_5445_axes_0 = const()[name = string("op_5445_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_5445_cast_fp16 = squeeze(axes = var_5445_axes_0, x = V_n_17_cast_fp16)[name = string("op_5445_cast_fp16")];
+            tensor<int32, [4]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_18_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_18_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_18_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_66, begin_mask = kv_cache_0_internal_tensor_assign_18_begin_mask_0, end = concat_67, end_mask = kv_cache_0_internal_tensor_assign_18_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_18_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_18_stride_0, update = var_5445_cast_fp16, x = coreml_update_state_40)[name = string("kv_cache_0_internal_tensor_assign_18_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_18_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_41_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_41 = read_state(input = kv_cache_0)[name = string("coreml_update_state_41")];
+            tensor<int32, [4]> var_5455_begin_0 = const()[name = string("op_5455_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_5455_end_0 = const()[name = string("op_5455_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_5455_end_mask_0 = const()[name = string("op_5455_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_5455_cast_fp16 = slice_by_index(begin = var_5455_begin_0, end = var_5455_end_0, end_mask = var_5455_end_mask_0, x = K_n_17_cast_fp16)[name = string("op_5455_cast_fp16")];
+            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = var_5455_cast_fp16)[name = string("transpose_55")];
+            tensor<fp16, [8, 1, 512, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
+            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_32_cast_fp16 = reshape(shape = concat_68, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
+            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_33_cast_fp16 = reshape(shape = concat_69, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
+            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_5464_begin_0 = const()[name = string("op_5464_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_5464_end_0 = const()[name = string("op_5464_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_5464_end_mask_0 = const()[name = string("op_5464_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_5464_cast_fp16 = slice_by_index(begin = var_5464_begin_0, end = var_5464_end_0, end_mask = var_5464_end_mask_0, x = V_n_17_cast_fp16)[name = string("op_5464_cast_fp16")];
+            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = var_5464_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [8, 1, 512, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
+            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_34_cast_fp16 = reshape(shape = concat_70, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
+            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_35_cast_fp16 = reshape(shape = concat_71, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
+            tensor<int32, [4]> Ve_17_perm_0 = const()[name = string("Ve_17_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_5482_transpose_x_0 = const()[name = string("op_5482_transpose_x_0"), val = bool(false)];
+            bool var_5482_transpose_y_0 = const()[name = string("op_5482_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [1, 8, 1, 512]> var_5482_cast_fp16 = matmul(transpose_x = var_5482_transpose_x_0, transpose_y = var_5482_transpose_y_0, x = q_71, y = transpose_56_cast_fp16)[name = string("op_5482_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_5489_cast_fp16 = add(x = var_5482_cast_fp16, y = causal_mask)[name = string("op_5489_cast_fp16")];
+            int32 var_5490 = const()[name = string("op_5490"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_5492_cast_fp16 = softmax(axis = var_5490, x = var_5489_cast_fp16)[name = string("op_5492_cast_fp16")];
+            bool var_5508_transpose_x_0 = const()[name = string("op_5508_transpose_x_0"), val = bool(false)];
+            bool var_5508_transpose_y_0 = const()[name = string("op_5508_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_17_cast_fp16 = transpose(perm = Ve_17_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [1, 8, 1, 256]> var_5508_cast_fp16 = matmul(transpose_x = var_5508_transpose_x_0, transpose_y = var_5508_transpose_y_0, x = var_5492_cast_fp16, y = Ve_17_cast_fp16)[name = string("op_5508_cast_fp16")];
+            tensor<int32, [4]> var_5518 = const()[name = string("op_5518"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5525 = const()[name = string("op_5525"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_5519 = transpose(perm = var_5518, x = var_5508_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 1, 2048]> var_5526 = reshape(shape = var_5525, x = var_5519)[name = string("op_5526")];
+            tensor<int32, [3]> var_5530 = const()[name = string("op_5530"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_8_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216657600))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218230528))))[name = string("squeeze_8_palettized")];
+            string var_5546_pad_type_0 = const()[name = string("op_5546_pad_type_0"), val = string("valid")];
+            int32 var_5546_groups_0 = const()[name = string("op_5546_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_5546_strides_0 = const()[name = string("op_5546_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_5546_pad_0 = const()[name = string("op_5546_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_5546_dilations_0 = const()[name = string("op_5546_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_5531 = transpose(perm = var_5530, x = var_5526)[name = string("transpose_48")];
+            tensor<fp16, [1, 1536, 1]> var_5546 = conv(dilations = var_5546_dilations_0, groups = var_5546_groups_0, pad = var_5546_pad_0, pad_type = var_5546_pad_type_0, strides = var_5546_strides_0, weight = squeeze_8_palettized, x = var_5531)[name = string("op_5546")];
+            tensor<int32, [3]> var_5550 = const()[name = string("op_5550"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_5556 = const()[name = string("op_5556"), val = int32(-1)];
+            fp16 const_152_promoted_to_fp16 = const()[name = string("const_152_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_253 = transpose(perm = var_5550, x = var_5546)[name = string("transpose_47")];
+            tensor<fp16, [1, 1, 1536]> var_5562_cast_fp16 = mul(x = x_253, y = const_152_promoted_to_fp16)[name = string("op_5562_cast_fp16")];
+            bool input_251_interleave_0 = const()[name = string("input_251_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_251_cast_fp16 = concat(axis = var_5556, interleave = input_251_interleave_0, values = (x_253, var_5562_cast_fp16))[name = string("input_251_cast_fp16")];
+            tensor<int32, [1]> normed_237_axes_0 = const()[name = string("normed_237_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_237_cast_fp16 = layer_norm(axes = normed_237_axes_0, epsilon = var_5554_to_fp16, x = input_251_cast_fp16)[name = string("normed_237_cast_fp16")];
+            tensor<int32, [2]> var_5567_split_sizes_0 = const()[name = string("op_5567_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5567_axis_0 = const()[name = string("op_5567_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5567_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5567_cast_fp16_1 = split(axis = var_5567_axis_0, split_sizes = var_5567_split_sizes_0, x = normed_237_cast_fp16)[name = string("op_5567_cast_fp16")];
+            tensor<fp16, [1536]> const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218232128)))];
+            tensor<fp16, [1, 1, 1536]> var_5570_cast_fp16 = mul(x = var_5567_cast_fp16_0, y = const_153_to_fp16)[name = string("op_5570_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_257_cast_fp16 = add(x = x_239_cast_fp16, y = var_5570_cast_fp16)[name = string("x_257_cast_fp16")];
+            int32 var_5577 = const()[name = string("op_5577"), 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_5583_cast_fp16 = mul(x = x_257_cast_fp16, y = const_154_promoted_to_fp16)[name = string("op_5583_cast_fp16")];
+            bool input_253_interleave_0 = const()[name = string("input_253_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_253_cast_fp16 = concat(axis = var_5577, interleave = input_253_interleave_0, values = (x_257_cast_fp16, var_5583_cast_fp16))[name = string("input_253_cast_fp16")];
+            tensor<int32, [1]> normed_241_axes_0 = const()[name = string("normed_241_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5575_to_fp16 = const()[name = string("op_5575_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_241_cast_fp16 = layer_norm(axes = normed_241_axes_0, epsilon = var_5575_to_fp16, x = input_253_cast_fp16)[name = string("normed_241_cast_fp16")];
+            tensor<int32, [2]> var_5588_split_sizes_0 = const()[name = string("op_5588_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5588_axis_0 = const()[name = string("op_5588_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5588_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5588_cast_fp16_1 = split(axis = var_5588_axis_0, split_sizes = var_5588_split_sizes_0, x = normed_241_cast_fp16)[name = string("op_5588_cast_fp16")];
+            tensor<fp16, [1536]> const_155_to_fp16 = const()[name = string("const_155_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218235264)))];
+            tensor<fp16, [1, 1, 1536]> var_5591_cast_fp16 = mul(x = var_5588_cast_fp16_0, y = const_155_to_fp16)[name = string("op_5591_cast_fp16")];
+            tensor<int32, [3]> var_5604 = const()[name = string("op_5604"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_255_axes_0 = const()[name = string("input_255_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_5605 = transpose(perm = var_5604, x = var_5591_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [1, 1536, 1, 1]> input_255 = expand_dims(axes = input_255_axes_0, x = var_5605)[name = string("input_255")];
+            string var_5618_pad_type_0 = const()[name = string("op_5618_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5618_strides_0 = const()[name = string("op_5618_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5618_pad_0 = const()[name = string("op_5618_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5618_dilations_0 = const()[name = string("op_5618_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5618_groups_0 = const()[name = string("op_5618_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_5618 = conv(dilations = var_5618_dilations_0, groups = var_5618_groups_0, pad = var_5618_pad_0, pad_type = var_5618_pad_type_0, strides = var_5618_strides_0, weight = layers_8_mlp_gate_proj_weight_palettized, x = input_255)[name = string("op_5618")];
+            string var_5620_mode_0 = const()[name = string("op_5620_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_5620 = gelu(mode = var_5620_mode_0, x = var_5618)[name = string("op_5620")];
+            string var_5631_pad_type_0 = const()[name = string("op_5631_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5631_strides_0 = const()[name = string("op_5631_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5631_pad_0 = const()[name = string("op_5631_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5631_dilations_0 = const()[name = string("op_5631_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5631_groups_0 = const()[name = string("op_5631_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_5631 = conv(dilations = var_5631_dilations_0, groups = var_5631_groups_0, pad = var_5631_pad_0, pad_type = var_5631_pad_type_0, strides = var_5631_strides_0, weight = layers_8_mlp_up_proj_weight_palettized, x = input_255)[name = string("op_5631")];
+            tensor<fp16, [1, 6144, 1, 1]> input_257 = mul(x = var_5620, y = var_5631)[name = string("input_257")];
+            string var_5643_pad_type_0 = const()[name = string("op_5643_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5643_strides_0 = const()[name = string("op_5643_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5643_pad_0 = const()[name = string("op_5643_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5643_dilations_0 = const()[name = string("op_5643_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5643_groups_0 = const()[name = string("op_5643_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_5643 = conv(dilations = var_5643_dilations_0, groups = var_5643_groups_0, pad = var_5643_pad_0, pad_type = var_5643_pad_type_0, strides = var_5643_strides_0, weight = layers_8_mlp_down_proj_weight_palettized, x = input_257)[name = string("op_5643")];
+            tensor<int32, [1]> var_5645_axes_0 = const()[name = string("op_5645_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_5645 = squeeze(axes = var_5645_axes_0, x = var_5643)[name = string("op_5645")];
+            tensor<int32, [3]> var_5649 = const()[name = string("op_5649"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_5655 = const()[name = string("op_5655"), val = int32(-1)];
+            fp16 const_156_promoted_to_fp16 = const()[name = string("const_156_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_261 = transpose(perm = var_5649, x = var_5645)[name = string("transpose_45")];
+            tensor<fp16, [1, 1, 1536]> var_5661_cast_fp16 = mul(x = x_261, y = const_156_promoted_to_fp16)[name = string("op_5661_cast_fp16")];
+            bool input_259_interleave_0 = const()[name = string("input_259_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_259_cast_fp16 = concat(axis = var_5655, interleave = input_259_interleave_0, values = (x_261, var_5661_cast_fp16))[name = string("input_259_cast_fp16")];
+            tensor<int32, [1]> normed_245_axes_0 = const()[name = string("normed_245_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5653_to_fp16 = const()[name = string("op_5653_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_245_cast_fp16 = layer_norm(axes = normed_245_axes_0, epsilon = var_5653_to_fp16, x = input_259_cast_fp16)[name = string("normed_245_cast_fp16")];
+            tensor<int32, [2]> var_5666_split_sizes_0 = const()[name = string("op_5666_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5666_axis_0 = const()[name = string("op_5666_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5666_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5666_cast_fp16_1 = split(axis = var_5666_axis_0, split_sizes = var_5666_split_sizes_0, x = normed_245_cast_fp16)[name = string("op_5666_cast_fp16")];
+            tensor<fp16, [1536]> const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218238400)))];
+            tensor<fp16, [1, 1, 1536]> var_5669_cast_fp16 = mul(x = var_5666_cast_fp16_0, y = const_157_to_fp16)[name = string("op_5669_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_125_cast_fp16 = add(x = x_257_cast_fp16, y = var_5669_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_5680 = linear(bias = linear_0_bias_0, weight = layers_8_per_layer_input_gate_weight_palettized, x = hidden_states_125_cast_fp16)[name = string("linear_16")];
+            string gated_17_mode_0 = const()[name = string("gated_17_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_17 = gelu(mode = gated_17_mode_0, x = var_5680)[name = string("gated_17")];
+            tensor<int32, [3]> var_5697_begin_0 = const()[name = string("op_5697_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> var_5697_end_0 = const()[name = string("op_5697_end_0"), val = tensor<int32, [3]>([1, 1, 2304])];
+            tensor<bool, [3]> var_5697_end_mask_0 = const()[name = string("op_5697_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_5697_cast_fp16 = slice_by_index(begin = var_5697_begin_0, end = var_5697_end_0, end_mask = var_5697_end_mask_0, x = per_layer_combined)[name = string("op_5697_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_263_cast_fp16 = mul(x = gated_17, y = var_5697_cast_fp16)[name = string("input_263_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_8_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218241536))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218438208))))[name = string("layers_8_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_263_cast_fp16)[name = string("linear_17_cast_fp16")];
+            int32 var_5706 = const()[name = string("op_5706"), val = int32(-1)];
+            fp16 const_158_promoted_to_fp16 = const()[name = string("const_158_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_5712_cast_fp16 = mul(x = linear_17_cast_fp16, y = const_158_promoted_to_fp16)[name = string("op_5712_cast_fp16")];
+            bool input_265_interleave_0 = const()[name = string("input_265_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_265_cast_fp16 = concat(axis = var_5706, interleave = input_265_interleave_0, values = (linear_17_cast_fp16, var_5712_cast_fp16))[name = string("input_265_cast_fp16")];
+            tensor<int32, [1]> normed_249_axes_0 = const()[name = string("normed_249_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5704_to_fp16 = const()[name = string("op_5704_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_249_cast_fp16 = layer_norm(axes = normed_249_axes_0, epsilon = var_5704_to_fp16, x = input_265_cast_fp16)[name = string("normed_249_cast_fp16")];
+            tensor<int32, [2]> var_5717_split_sizes_0 = const()[name = string("op_5717_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5717_axis_0 = const()[name = string("op_5717_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5717_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5717_cast_fp16_1 = split(axis = var_5717_axis_0, split_sizes = var_5717_split_sizes_0, x = normed_249_cast_fp16)[name = string("op_5717_cast_fp16")];
+            tensor<fp16, [1536]> const_159_to_fp16 = const()[name = string("const_159_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218439808)))];
+            tensor<fp16, [1, 1, 1536]> var_5720_cast_fp16 = mul(x = var_5717_cast_fp16_0, y = const_159_to_fp16)[name = string("op_5720_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = var_5720_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
+            tensor<fp16, [1]> layers_8_layer_scalar_to_fp16 = const()[name = string("layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.82p-2])];
+            tensor<fp16, [1, 1, 1536]> x_269_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = layers_8_layer_scalar_to_fp16)[name = string("x_269_cast_fp16")];
+            int32 var_5728 = const()[name = string("op_5728"), val = int32(-1)];
+            fp16 const_160_promoted_to_fp16 = const()[name = string("const_160_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_5734_cast_fp16 = mul(x = x_269_cast_fp16, y = const_160_promoted_to_fp16)[name = string("op_5734_cast_fp16")];
+            bool input_267_interleave_0 = const()[name = string("input_267_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_267_cast_fp16 = concat(axis = var_5728, interleave = input_267_interleave_0, values = (x_269_cast_fp16, var_5734_cast_fp16))[name = string("input_267_cast_fp16")];
+            tensor<int32, [1]> normed_253_axes_0 = const()[name = string("normed_253_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5726_to_fp16 = const()[name = string("op_5726_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_253_cast_fp16 = layer_norm(axes = normed_253_axes_0, epsilon = var_5726_to_fp16, x = input_267_cast_fp16)[name = string("normed_253_cast_fp16")];
+            tensor<int32, [2]> var_5739_split_sizes_0 = const()[name = string("op_5739_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_5739_axis_0 = const()[name = string("op_5739_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_5739_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_5739_cast_fp16_1 = split(axis = var_5739_axis_0, split_sizes = var_5739_split_sizes_0, x = normed_253_cast_fp16)[name = string("op_5739_cast_fp16")];
+            tensor<fp16, [1536]> const_161_to_fp16 = const()[name = string("const_161_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218442944)))];
+            tensor<fp16, [1, 1, 1536]> var_5742_cast_fp16 = mul(x = var_5739_cast_fp16_0, y = const_161_to_fp16)[name = string("op_5742_cast_fp16")];
+            tensor<int32, [3]> var_5750 = const()[name = string("op_5750"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_5753_axes_0 = const()[name = string("op_5753_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_5751_cast_fp16 = transpose(perm = var_5750, x = var_5742_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [1, 1536, 1, 1]> var_5753_cast_fp16 = expand_dims(axes = var_5753_axes_0, x = var_5751_cast_fp16)[name = string("op_5753_cast_fp16")];
+            string var_5769_pad_type_0 = const()[name = string("op_5769_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5769_strides_0 = const()[name = string("op_5769_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5769_pad_0 = const()[name = string("op_5769_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5769_dilations_0 = const()[name = string("op_5769_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5769_groups_0 = const()[name = string("op_5769_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 4096, 1, 1]> var_5769 = conv(dilations = var_5769_dilations_0, groups = var_5769_groups_0, pad = var_5769_pad_0, pad_type = var_5769_pad_type_0, strides = var_5769_strides_0, weight = layers_9_self_attn_q_proj_weight_palettized, x = var_5753_cast_fp16)[name = string("op_5769")];
+            tensor<int32, [4]> var_5774 = const()[name = string("op_5774"), val = tensor<int32, [4]>([1, 8, 512, 1])];
+            tensor<fp16, [1, 8, 512, 1]> var_5775 = reshape(shape = var_5774, x = var_5769)[name = string("op_5775")];
+            tensor<int32, [4]> var_5780 = const()[name = string("op_5780"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_5790 = const()[name = string("op_5790"), val = tensor<int32, [3]>([1, 8, 512])];
+            tensor<fp16, [1, 8, 1, 512]> var_5781 = transpose(perm = var_5780, x = var_5775)[name = string("transpose_43")];
+            tensor<fp16, [1, 8, 512]> x_273 = reshape(shape = var_5790, x = var_5781)[name = string("x_273")];
+            int32 var_5796 = const()[name = string("op_5796"), val = int32(-1)];
+            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 512]> var_5802_cast_fp16 = mul(x = x_273, y = const_162_promoted_to_fp16)[name = string("op_5802_cast_fp16")];
+            bool input_271_interleave_0 = const()[name = string("input_271_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1024]> input_271_cast_fp16 = concat(axis = var_5796, interleave = input_271_interleave_0, values = (x_273, var_5802_cast_fp16))[name = string("input_271_cast_fp16")];
+            tensor<int32, [1]> normed_257_axes_0 = const()[name = string("normed_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5794_to_fp16 = const()[name = string("op_5794_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 1024]> normed_257_cast_fp16 = layer_norm(axes = normed_257_axes_0, epsilon = var_5794_to_fp16, x = input_271_cast_fp16)[name = string("normed_257_cast_fp16")];
+            tensor<int32, [2]> var_5807_split_sizes_0 = const()[name = string("op_5807_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            int32 var_5807_axis_0 = const()[name = string("op_5807_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 512]> var_5807_cast_fp16_0, tensor<fp16, [1, 8, 512]> var_5807_cast_fp16_1 = split(axis = var_5807_axis_0, split_sizes = var_5807_split_sizes_0, x = normed_257_cast_fp16)[name = string("op_5807_cast_fp16")];
+            tensor<fp16, [512]> const_163_to_fp16 = const()[name = string("const_163_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218446080)))];
+            tensor<fp16, [1, 8, 512]> var_5810_cast_fp16 = mul(x = var_5807_cast_fp16_0, y = const_163_to_fp16)[name = string("op_5810_cast_fp16")];
+            tensor<int32, [4]> var_5816 = const()[name = string("op_5816"), val = tensor<int32, [4]>([1, 8, 1, 512])];
+            tensor<fp16, [1, 8, 1, 512]> q_75 = reshape(shape = var_5816, x = var_5810_cast_fp16)[name = string("q_75")];
+            tensor<fp16, [1, 8, 1, 512]> var_5818 = mul(x = q_75, y = cos)[name = string("op_5818")];
+            tensor<int32, [2]> var_5819_split_sizes_0 = const()[name = string("op_5819_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_5819_axis_0 = const()[name = string("op_5819_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 256]> var_5819_0, tensor<fp16, [1, 8, 1, 256]> var_5819_1 = split(axis = var_5819_axis_0, split_sizes = var_5819_split_sizes_0, x = q_75)[name = string("op_5819")];
+            fp16 const_164_promoted = const()[name = string("const_164_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 256]> var_5821 = mul(x = var_5819_1, y = const_164_promoted)[name = string("op_5821")];
+            int32 var_5823 = const()[name = string("op_5823"), val = int32(-1)];
+            bool var_5824_interleave_0 = const()[name = string("op_5824_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 512]> var_5824 = concat(axis = var_5823, interleave = var_5824_interleave_0, values = (var_5821, var_5819_0))[name = string("op_5824")];
+            tensor<fp16, [1, 8, 1, 512]> var_5825 = mul(x = var_5824, y = sin)[name = string("op_5825")];
+            tensor<fp16, [1, 8, 1, 512]> q_79 = add(x = var_5818, y = var_5825)[name = string("q_79")];
+            string var_5838_pad_type_0 = const()[name = string("op_5838_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5838_strides_0 = const()[name = string("op_5838_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5838_pad_0 = const()[name = string("op_5838_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5838_dilations_0 = const()[name = string("op_5838_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5838_groups_0 = const()[name = string("op_5838_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 512, 1, 1]> var_5838 = conv(dilations = var_5838_dilations_0, groups = var_5838_groups_0, pad = var_5838_pad_0, pad_type = var_5838_pad_type_0, strides = var_5838_strides_0, weight = layers_9_self_attn_k_proj_weight_palettized, x = var_5753_cast_fp16)[name = string("op_5838")];
+            tensor<int32, [4]> var_5843 = const()[name = string("op_5843"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 1]> var_5844 = reshape(shape = var_5843, x = var_5838)[name = string("op_5844")];
+            tensor<int32, [4]> var_5849 = const()[name = string("op_5849"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_5866_pad_type_0 = const()[name = string("op_5866_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_5866_strides_0 = const()[name = string("op_5866_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_5866_pad_0 = const()[name = string("op_5866_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_5866_dilations_0 = const()[name = string("op_5866_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_5866_groups_0 = const()[name = string("op_5866_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 512, 1, 1]> var_5866 = conv(dilations = var_5866_dilations_0, groups = var_5866_groups_0, pad = var_5866_pad_0, pad_type = var_5866_pad_type_0, strides = var_5866_strides_0, weight = layers_9_self_attn_v_proj_weight_palettized, x = var_5753_cast_fp16)[name = string("op_5866")];
+            tensor<int32, [4]> var_5871 = const()[name = string("op_5871"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 1]> var_5872 = reshape(shape = var_5871, x = var_5866)[name = string("op_5872")];
+            tensor<int32, [4]> var_5877 = const()[name = string("op_5877"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_5887 = const()[name = string("op_5887"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<fp16, [1, 1, 1, 512]> var_5850 = transpose(perm = var_5849, x = var_5844)[name = string("transpose_42")];
+            tensor<fp16, [1, 1, 512]> x_277 = reshape(shape = var_5887, x = var_5850)[name = string("x_277")];
+            int32 var_5893 = const()[name = string("op_5893"), val = int32(-1)];
+            fp16 const_165_promoted_to_fp16 = const()[name = string("const_165_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 512]> var_5899_cast_fp16 = mul(x = x_277, y = const_165_promoted_to_fp16)[name = string("op_5899_cast_fp16")];
+            bool input_273_interleave_0 = const()[name = string("input_273_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1024]> input_273_cast_fp16 = concat(axis = var_5893, interleave = input_273_interleave_0, values = (x_277, var_5899_cast_fp16))[name = string("input_273_cast_fp16")];
+            tensor<int32, [1]> normed_261_axes_0 = const()[name = string("normed_261_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_5891_to_fp16 = const()[name = string("op_5891_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1024]> normed_261_cast_fp16 = layer_norm(axes = normed_261_axes_0, epsilon = var_5891_to_fp16, x = input_273_cast_fp16)[name = string("normed_261_cast_fp16")];
+            tensor<int32, [2]> var_5904_split_sizes_0 = const()[name = string("op_5904_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            int32 var_5904_axis_0 = const()[name = string("op_5904_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 512]> var_5904_cast_fp16_0, tensor<fp16, [1, 1, 512]> var_5904_cast_fp16_1 = split(axis = var_5904_axis_0, split_sizes = var_5904_split_sizes_0, x = normed_261_cast_fp16)[name = string("op_5904_cast_fp16")];
+            tensor<fp16, [512]> const_166_to_fp16 = const()[name = string("const_166_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218447168)))];
+            tensor<fp16, [1, 1, 512]> var_5907_cast_fp16 = mul(x = var_5904_cast_fp16_0, y = const_166_to_fp16)[name = string("op_5907_cast_fp16")];
+            tensor<int32, [4]> var_5913 = const()[name = string("op_5913"), val = tensor<int32, [4]>([1, 1, 1, 512])];
+            tensor<fp16, [1, 1, 1, 512]> q_77 = reshape(shape = var_5913, x = var_5907_cast_fp16)[name = string("q_77")];
+            fp16 var_5920_promoted_to_fp16 = const()[name = string("op_5920_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 512]> var_5878 = transpose(perm = var_5877, x = var_5872)[name = string("transpose_41")];
+            tensor<fp16, [1, 1, 1, 512]> var_5921_cast_fp16 = pow(x = var_5878, y = var_5920_promoted_to_fp16)[name = string("op_5921_cast_fp16")];
+            tensor<int32, [1]> var_5926_axes_0 = const()[name = string("op_5926_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_5926_keep_dims_0 = const()[name = string("op_5926_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_5926_cast_fp16 = reduce_mean(axes = var_5926_axes_0, keep_dims = var_5926_keep_dims_0, x = var_5921_cast_fp16)[name = string("op_5926_cast_fp16")];
+            fp16 var_5928_to_fp16 = const()[name = string("op_5928_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_19_cast_fp16 = add(x = var_5926_cast_fp16, y = var_5928_to_fp16)[name = string("mean_sq_19_cast_fp16")];
+            fp16 var_5935_to_fp16 = const()[name = string("op_5935_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_5936_cast_fp16 = pow(x = mean_sq_19_cast_fp16, y = var_5935_to_fp16)[name = string("op_5936_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 512]> var_5937_cast_fp16 = mul(x = var_5878, y = var_5936_cast_fp16)[name = string("op_5937_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 512]> var_5943 = mul(x = q_77, y = cos)[name = string("op_5943")];
+            tensor<int32, [2]> var_5944_split_sizes_0 = const()[name = string("op_5944_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_5944_axis_0 = const()[name = string("op_5944_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 256]> var_5944_0, tensor<fp16, [1, 1, 1, 256]> var_5944_1 = split(axis = var_5944_axis_0, split_sizes = var_5944_split_sizes_0, x = q_77)[name = string("op_5944")];
+            fp16 const_167_promoted = const()[name = string("const_167_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 256]> var_5946 = mul(x = var_5944_1, y = const_167_promoted)[name = string("op_5946")];
+            int32 var_5948 = const()[name = string("op_5948"), val = int32(-1)];
+            bool var_5949_interleave_0 = const()[name = string("op_5949_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 512]> var_5949 = concat(axis = var_5948, interleave = var_5949_interleave_0, values = (var_5946, var_5944_0))[name = string("op_5949")];
+            tensor<fp16, [1, 1, 1, 512]> var_5950 = mul(x = var_5949, y = sin)[name = string("op_5950")];
+            tensor<fp16, [1, 1, 1, 512]> k_23 = add(x = var_5943, y = var_5950)[name = string("k_23")];
+            tensor<int32, [4]> var_5955_begin_0 = const()[name = string("op_5955_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> var_5955_end_0 = const()[name = string("op_5955_end_0"), val = tensor<int32, [4]>([10, 1, 512, 512])];
+            tensor<bool, [4]> var_5955_end_mask_0 = const()[name = string("op_5955_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_5955_squeeze_mask_0 = const()[name = string("op_5955_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_5955_cast_fp16 = slice_by_index(begin = var_5955_begin_0, end = var_5955_end_0, end_mask = var_5955_end_mask_0, squeeze_mask = var_5955_squeeze_mask_0, x = coreml_update_state_41)[name = string("op_5955_cast_fp16")];
+            tensor<int32, [1]> K_c_19_axes_0 = const()[name = string("K_c_19_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_19_cast_fp16 = expand_dims(axes = K_c_19_axes_0, x = var_5955_cast_fp16)[name = string("K_c_19_cast_fp16")];
+            tensor<int32, [4]> var_5960_begin_0 = const()[name = string("op_5960_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> var_5960_end_0 = const()[name = string("op_5960_end_0"), val = tensor<int32, [4]>([22, 1, 512, 512])];
+            tensor<bool, [4]> var_5960_end_mask_0 = const()[name = string("op_5960_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_5960_squeeze_mask_0 = const()[name = string("op_5960_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_5960_cast_fp16 = slice_by_index(begin = var_5960_begin_0, end = var_5960_end_0, end_mask = var_5960_end_mask_0, squeeze_mask = var_5960_squeeze_mask_0, x = coreml_update_state_41)[name = string("op_5960_cast_fp16")];
+            tensor<int32, [1]> V_c_19_axes_0 = const()[name = string("V_c_19_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_19_cast_fp16 = expand_dims(axes = V_c_19_axes_0, x = var_5960_cast_fp16)[name = string("V_c_19_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_5966_cast_fp16 = mul(x = K_c_19_cast_fp16, y = var_1005_cast_fp16)[name = string("op_5966_cast_fp16")];
+            tensor<int32, [4]> var_5967_reps_0 = const()[name = string("op_5967_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_5967 = tile(reps = var_5967_reps_0, x = k_23)[name = string("op_5967")];
+            tensor<fp16, [1, 1, 512, 512]> var_5968_cast_fp16 = mul(x = var_5967, y = update_mask)[name = string("op_5968_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_19_cast_fp16 = add(x = var_5966_cast_fp16, y = var_5968_cast_fp16)[name = string("K_n_19_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_5974_cast_fp16 = mul(x = V_c_19_cast_fp16, y = var_1005_cast_fp16)[name = string("op_5974_cast_fp16")];
+            tensor<int32, [4]> var_5975_reps_0 = const()[name = string("op_5975_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_5975 = tile(reps = var_5975_reps_0, x = var_5937_cast_fp16)[name = string("op_5975")];
+            tensor<fp16, [1, 1, 512, 512]> var_5976_cast_fp16 = mul(x = var_5975, y = update_mask)[name = string("op_5976_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_19_cast_fp16 = add(x = var_5974_cast_fp16, y = var_5976_cast_fp16)[name = string("V_n_19_cast_fp16")];
+            tensor<int32, [1]> var_5980_axes_0 = const()[name = string("op_5980_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_5980_cast_fp16 = squeeze(axes = var_5980_axes_0, x = K_n_19_cast_fp16)[name = string("op_5980_cast_fp16")];
+            tensor<int32, [4]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_19_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_19_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_19_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_72, begin_mask = kv_cache_0_internal_tensor_assign_19_begin_mask_0, end = concat_73, end_mask = kv_cache_0_internal_tensor_assign_19_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_19_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_19_stride_0, update = var_5980_cast_fp16, x = coreml_update_state_41)[name = string("kv_cache_0_internal_tensor_assign_19_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_19_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_42_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_42 = read_state(input = kv_cache_0)[name = string("coreml_update_state_42")];
+            tensor<int32, [1]> var_5987_axes_0 = const()[name = string("op_5987_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_5987_cast_fp16 = squeeze(axes = var_5987_axes_0, x = V_n_19_cast_fp16)[name = string("op_5987_cast_fp16")];
+            tensor<int32, [4]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_20_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_20_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_20_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_20_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_74, begin_mask = kv_cache_0_internal_tensor_assign_20_begin_mask_0, end = concat_75, end_mask = kv_cache_0_internal_tensor_assign_20_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_20_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_20_stride_0, update = var_5987_cast_fp16, x = coreml_update_state_42)[name = string("kv_cache_0_internal_tensor_assign_20_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_20_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_43_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_43 = read_state(input = kv_cache_0)[name = string("coreml_update_state_43")];
+            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 512]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = K_n_19_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [8, 1, 512, 512]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
+            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([8, 1, 1, 512, 512])];
+            tensor<fp16, [8, 1, 1, 512, 512]> reshape_36_cast_fp16 = reshape(shape = concat_76, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
+            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
+            tensor<fp16, [1, 8, 1, 512, 512]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_39")];
+            tensor<fp16, [8, 1, 512, 512]> reshape_37_cast_fp16 = reshape(shape = concat_77, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
+            tensor<int32, [4]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 512]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = V_n_19_cast_fp16)[name = string("transpose_38")];
+            tensor<fp16, [8, 1, 512, 512]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
+            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([8, 1, 1, 512, 512])];
+            tensor<fp16, [8, 1, 1, 512, 512]> reshape_38_cast_fp16 = reshape(shape = concat_78, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
+            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
+            tensor<fp16, [1, 8, 1, 512, 512]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_37")];
+            tensor<fp16, [8, 1, 512, 512]> reshape_39_cast_fp16 = reshape(shape = concat_79, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
+            tensor<int32, [4]> Ve_19_perm_0 = const()[name = string("Ve_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_6024_transpose_x_0 = const()[name = string("op_6024_transpose_x_0"), val = bool(false)];
+            bool var_6024_transpose_y_0 = const()[name = string("op_6024_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 512]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_36")];
+            tensor<fp16, [1, 8, 1, 512]> var_6024_cast_fp16 = matmul(transpose_x = var_6024_transpose_x_0, transpose_y = var_6024_transpose_y_0, x = q_79, y = transpose_57_cast_fp16)[name = string("op_6024_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_6031_cast_fp16 = add(x = var_6024_cast_fp16, y = causal_mask)[name = string("op_6031_cast_fp16")];
+            int32 var_6032 = const()[name = string("op_6032"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_6034_cast_fp16 = softmax(axis = var_6032, x = var_6031_cast_fp16)[name = string("op_6034_cast_fp16")];
+            bool var_6050_transpose_x_0 = const()[name = string("op_6050_transpose_x_0"), val = bool(false)];
+            bool var_6050_transpose_y_0 = const()[name = string("op_6050_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 512]> Ve_19_cast_fp16 = transpose(perm = Ve_19_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_35")];
+            tensor<fp16, [1, 8, 1, 512]> var_6050_cast_fp16 = matmul(transpose_x = var_6050_transpose_x_0, transpose_y = var_6050_transpose_y_0, x = var_6034_cast_fp16, y = Ve_19_cast_fp16)[name = string("op_6050_cast_fp16")];
+            tensor<int32, [4]> var_6060 = const()[name = string("op_6060"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_6067 = const()[name = string("op_6067"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 512]> var_6061 = transpose(perm = var_6060, x = var_6050_cast_fp16)[name = string("transpose_34")];
+            tensor<fp16, [1, 1, 4096]> var_6068 = reshape(shape = var_6067, x = var_6061)[name = string("op_6068")];
+            tensor<int32, [3]> var_6072 = const()[name = string("op_6072"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 4096, 1]> squeeze_9_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218448256))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221594048))))[name = string("squeeze_9_palettized")];
+            string var_6088_pad_type_0 = const()[name = string("op_6088_pad_type_0"), val = string("valid")];
+            int32 var_6088_groups_0 = const()[name = string("op_6088_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_6088_strides_0 = const()[name = string("op_6088_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_6088_pad_0 = const()[name = string("op_6088_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_6088_dilations_0 = const()[name = string("op_6088_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 4096, 1]> var_6073 = transpose(perm = var_6072, x = var_6068)[name = string("transpose_33")];
+            tensor<fp16, [1, 1536, 1]> var_6088 = conv(dilations = var_6088_dilations_0, groups = var_6088_groups_0, pad = var_6088_pad_0, pad_type = var_6088_pad_type_0, strides = var_6088_strides_0, weight = squeeze_9_palettized, x = var_6073)[name = string("op_6088")];
+            tensor<int32, [3]> var_6092 = const()[name = string("op_6092"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_6098 = const()[name = string("op_6098"), val = int32(-1)];
+            fp16 const_168_promoted_to_fp16 = const()[name = string("const_168_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_283 = transpose(perm = var_6092, x = var_6088)[name = string("transpose_32")];
+            tensor<fp16, [1, 1, 1536]> var_6104_cast_fp16 = mul(x = x_283, y = const_168_promoted_to_fp16)[name = string("op_6104_cast_fp16")];
+            bool input_277_interleave_0 = const()[name = string("input_277_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_277_cast_fp16 = concat(axis = var_6098, interleave = input_277_interleave_0, values = (x_283, var_6104_cast_fp16))[name = string("input_277_cast_fp16")];
+            tensor<int32, [1]> normed_265_axes_0 = const()[name = string("normed_265_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6096_to_fp16 = const()[name = string("op_6096_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_265_cast_fp16 = layer_norm(axes = normed_265_axes_0, epsilon = var_6096_to_fp16, x = input_277_cast_fp16)[name = string("normed_265_cast_fp16")];
+            tensor<int32, [2]> var_6109_split_sizes_0 = const()[name = string("op_6109_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6109_axis_0 = const()[name = string("op_6109_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6109_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6109_cast_fp16_1 = split(axis = var_6109_axis_0, split_sizes = var_6109_split_sizes_0, x = normed_265_cast_fp16)[name = string("op_6109_cast_fp16")];
+            tensor<fp16, [1536]> const_169_to_fp16 = const()[name = string("const_169_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221595648)))];
+            tensor<fp16, [1, 1, 1536]> var_6112_cast_fp16 = mul(x = var_6109_cast_fp16_0, y = const_169_to_fp16)[name = string("op_6112_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_287_cast_fp16 = add(x = x_269_cast_fp16, y = var_6112_cast_fp16)[name = string("x_287_cast_fp16")];
+            int32 var_6119 = const()[name = string("op_6119"), val = int32(-1)];
+            fp16 const_170_promoted_to_fp16 = const()[name = string("const_170_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_6125_cast_fp16 = mul(x = x_287_cast_fp16, y = const_170_promoted_to_fp16)[name = string("op_6125_cast_fp16")];
+            bool input_279_interleave_0 = const()[name = string("input_279_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_279_cast_fp16 = concat(axis = var_6119, interleave = input_279_interleave_0, values = (x_287_cast_fp16, var_6125_cast_fp16))[name = string("input_279_cast_fp16")];
+            tensor<int32, [1]> normed_269_axes_0 = const()[name = string("normed_269_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6117_to_fp16 = const()[name = string("op_6117_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_269_cast_fp16 = layer_norm(axes = normed_269_axes_0, epsilon = var_6117_to_fp16, x = input_279_cast_fp16)[name = string("normed_269_cast_fp16")];
+            tensor<int32, [2]> var_6130_split_sizes_0 = const()[name = string("op_6130_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6130_axis_0 = const()[name = string("op_6130_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6130_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6130_cast_fp16_1 = split(axis = var_6130_axis_0, split_sizes = var_6130_split_sizes_0, x = normed_269_cast_fp16)[name = string("op_6130_cast_fp16")];
+            tensor<fp16, [1536]> const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221598784)))];
+            tensor<fp16, [1, 1, 1536]> var_6133_cast_fp16 = mul(x = var_6130_cast_fp16_0, y = const_171_to_fp16)[name = string("op_6133_cast_fp16")];
+            tensor<int32, [3]> var_6146 = const()[name = string("op_6146"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_281_axes_0 = const()[name = string("input_281_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_6147 = transpose(perm = var_6146, x = var_6133_cast_fp16)[name = string("transpose_31")];
+            tensor<fp16, [1, 1536, 1, 1]> input_281 = expand_dims(axes = input_281_axes_0, x = var_6147)[name = string("input_281")];
+            string var_6160_pad_type_0 = const()[name = string("op_6160_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6160_strides_0 = const()[name = string("op_6160_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6160_pad_0 = const()[name = string("op_6160_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6160_dilations_0 = const()[name = string("op_6160_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6160_groups_0 = const()[name = string("op_6160_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_6160 = conv(dilations = var_6160_dilations_0, groups = var_6160_groups_0, pad = var_6160_pad_0, pad_type = var_6160_pad_type_0, strides = var_6160_strides_0, weight = layers_9_mlp_gate_proj_weight_palettized, x = input_281)[name = string("op_6160")];
+            string var_6162_mode_0 = const()[name = string("op_6162_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_6162 = gelu(mode = var_6162_mode_0, x = var_6160)[name = string("op_6162")];
+            string var_6173_pad_type_0 = const()[name = string("op_6173_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6173_strides_0 = const()[name = string("op_6173_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6173_pad_0 = const()[name = string("op_6173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6173_dilations_0 = const()[name = string("op_6173_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6173_groups_0 = const()[name = string("op_6173_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_6173 = conv(dilations = var_6173_dilations_0, groups = var_6173_groups_0, pad = var_6173_pad_0, pad_type = var_6173_pad_type_0, strides = var_6173_strides_0, weight = layers_9_mlp_up_proj_weight_palettized, x = input_281)[name = string("op_6173")];
+            tensor<fp16, [1, 6144, 1, 1]> input_283 = mul(x = var_6162, y = var_6173)[name = string("input_283")];
+            string var_6185_pad_type_0 = const()[name = string("op_6185_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6185_strides_0 = const()[name = string("op_6185_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6185_pad_0 = const()[name = string("op_6185_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6185_dilations_0 = const()[name = string("op_6185_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6185_groups_0 = const()[name = string("op_6185_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_6185 = conv(dilations = var_6185_dilations_0, groups = var_6185_groups_0, pad = var_6185_pad_0, pad_type = var_6185_pad_type_0, strides = var_6185_strides_0, weight = layers_9_mlp_down_proj_weight_palettized, x = input_283)[name = string("op_6185")];
+            tensor<int32, [1]> var_6187_axes_0 = const()[name = string("op_6187_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_6187 = squeeze(axes = var_6187_axes_0, x = var_6185)[name = string("op_6187")];
+            tensor<int32, [3]> var_6191 = const()[name = string("op_6191"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_6197 = const()[name = string("op_6197"), 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]> x_291 = transpose(perm = var_6191, x = var_6187)[name = string("transpose_30")];
+            tensor<fp16, [1, 1, 1536]> var_6203_cast_fp16 = mul(x = x_291, y = const_172_promoted_to_fp16)[name = string("op_6203_cast_fp16")];
+            bool input_285_interleave_0 = const()[name = string("input_285_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_285_cast_fp16 = concat(axis = var_6197, interleave = input_285_interleave_0, values = (x_291, var_6203_cast_fp16))[name = string("input_285_cast_fp16")];
+            tensor<int32, [1]> normed_273_axes_0 = const()[name = string("normed_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6195_to_fp16 = const()[name = string("op_6195_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_273_cast_fp16 = layer_norm(axes = normed_273_axes_0, epsilon = var_6195_to_fp16, x = input_285_cast_fp16)[name = string("normed_273_cast_fp16")];
+            tensor<int32, [2]> var_6208_split_sizes_0 = const()[name = string("op_6208_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6208_axis_0 = const()[name = string("op_6208_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6208_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6208_cast_fp16_1 = split(axis = var_6208_axis_0, split_sizes = var_6208_split_sizes_0, x = normed_273_cast_fp16)[name = string("op_6208_cast_fp16")];
+            tensor<fp16, [1536]> const_173_to_fp16 = const()[name = string("const_173_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221601920)))];
+            tensor<fp16, [1, 1, 1536]> var_6211_cast_fp16 = mul(x = var_6208_cast_fp16_0, y = const_173_to_fp16)[name = string("op_6211_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_139_cast_fp16 = add(x = x_287_cast_fp16, y = var_6211_cast_fp16)[name = string("hidden_states_139_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_6222 = linear(bias = linear_0_bias_0, weight = layers_9_per_layer_input_gate_weight_palettized, x = hidden_states_139_cast_fp16)[name = string("linear_18")];
+            string gated_19_mode_0 = const()[name = string("gated_19_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_19 = gelu(mode = gated_19_mode_0, x = var_6222)[name = string("gated_19")];
+            tensor<int32, [3]> var_6239_begin_0 = const()[name = string("op_6239_begin_0"), val = tensor<int32, [3]>([0, 0, 2304])];
+            tensor<int32, [3]> var_6239_end_0 = const()[name = string("op_6239_end_0"), val = tensor<int32, [3]>([1, 1, 2560])];
+            tensor<bool, [3]> var_6239_end_mask_0 = const()[name = string("op_6239_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_6239_cast_fp16 = slice_by_index(begin = var_6239_begin_0, end = var_6239_end_0, end_mask = var_6239_end_mask_0, x = per_layer_combined)[name = string("op_6239_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_289_cast_fp16 = mul(x = gated_19, y = var_6239_cast_fp16)[name = string("input_289_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_9_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221605056))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221801728))))[name = string("layers_9_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_289_cast_fp16)[name = string("linear_19_cast_fp16")];
+            int32 var_6248 = const()[name = string("op_6248"), val = int32(-1)];
+            fp16 const_174_promoted_to_fp16 = const()[name = string("const_174_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_6254_cast_fp16 = mul(x = linear_19_cast_fp16, y = const_174_promoted_to_fp16)[name = string("op_6254_cast_fp16")];
+            bool input_291_interleave_0 = const()[name = string("input_291_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_291_cast_fp16 = concat(axis = var_6248, interleave = input_291_interleave_0, values = (linear_19_cast_fp16, var_6254_cast_fp16))[name = string("input_291_cast_fp16")];
+            tensor<int32, [1]> normed_277_axes_0 = const()[name = string("normed_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6246_to_fp16 = const()[name = string("op_6246_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_277_cast_fp16 = layer_norm(axes = normed_277_axes_0, epsilon = var_6246_to_fp16, x = input_291_cast_fp16)[name = string("normed_277_cast_fp16")];
+            tensor<int32, [2]> var_6259_split_sizes_0 = const()[name = string("op_6259_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6259_axis_0 = const()[name = string("op_6259_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6259_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6259_cast_fp16_1 = split(axis = var_6259_axis_0, split_sizes = var_6259_split_sizes_0, x = normed_277_cast_fp16)[name = string("op_6259_cast_fp16")];
+            tensor<fp16, [1536]> const_175_to_fp16 = const()[name = string("const_175_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221803328)))];
+            tensor<fp16, [1, 1, 1536]> var_6262_cast_fp16 = mul(x = var_6259_cast_fp16_0, y = const_175_to_fp16)[name = string("op_6262_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_143_cast_fp16 = add(x = hidden_states_139_cast_fp16, y = var_6262_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
+            tensor<fp16, [1]> layers_9_layer_scalar_to_fp16 = const()[name = string("layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.dcp-2])];
+            tensor<fp16, [1, 1, 1536]> x_299_cast_fp16 = mul(x = hidden_states_143_cast_fp16, y = layers_9_layer_scalar_to_fp16)[name = string("x_299_cast_fp16")];
+            int32 var_6270 = const()[name = string("op_6270"), 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_6276_cast_fp16 = mul(x = x_299_cast_fp16, y = const_176_promoted_to_fp16)[name = string("op_6276_cast_fp16")];
+            bool input_293_interleave_0 = const()[name = string("input_293_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_293_cast_fp16 = concat(axis = var_6270, interleave = input_293_interleave_0, values = (x_299_cast_fp16, var_6276_cast_fp16))[name = string("input_293_cast_fp16")];
+            tensor<int32, [1]> normed_281_axes_0 = const()[name = string("normed_281_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6268_to_fp16 = const()[name = string("op_6268_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_281_cast_fp16 = layer_norm(axes = normed_281_axes_0, epsilon = var_6268_to_fp16, x = input_293_cast_fp16)[name = string("normed_281_cast_fp16")];
+            tensor<int32, [2]> var_6281_split_sizes_0 = const()[name = string("op_6281_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6281_axis_0 = const()[name = string("op_6281_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6281_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6281_cast_fp16_1 = split(axis = var_6281_axis_0, split_sizes = var_6281_split_sizes_0, x = normed_281_cast_fp16)[name = string("op_6281_cast_fp16")];
+            tensor<fp16, [1536]> const_177_to_fp16 = const()[name = string("const_177_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221806464)))];
+            tensor<fp16, [1, 1, 1536]> var_6284_cast_fp16 = mul(x = var_6281_cast_fp16_0, y = const_177_to_fp16)[name = string("op_6284_cast_fp16")];
+            tensor<int32, [3]> var_6292 = const()[name = string("op_6292"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_6295_axes_0 = const()[name = string("op_6295_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_6293_cast_fp16 = transpose(perm = var_6292, x = var_6284_cast_fp16)[name = string("transpose_29")];
+            tensor<fp16, [1, 1536, 1, 1]> var_6295_cast_fp16 = expand_dims(axes = var_6295_axes_0, x = var_6293_cast_fp16)[name = string("op_6295_cast_fp16")];
+            string var_6311_pad_type_0 = const()[name = string("op_6311_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6311_strides_0 = const()[name = string("op_6311_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6311_pad_0 = const()[name = string("op_6311_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6311_dilations_0 = const()[name = string("op_6311_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6311_groups_0 = const()[name = string("op_6311_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_6311 = conv(dilations = var_6311_dilations_0, groups = var_6311_groups_0, pad = var_6311_pad_0, pad_type = var_6311_pad_type_0, strides = var_6311_strides_0, weight = layers_10_self_attn_q_proj_weight_palettized, x = var_6295_cast_fp16)[name = string("op_6311")];
+            tensor<int32, [4]> var_6316 = const()[name = string("op_6316"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_6317 = reshape(shape = var_6316, x = var_6311)[name = string("op_6317")];
+            tensor<int32, [4]> var_6322 = const()[name = string("op_6322"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_6332 = const()[name = string("op_6332"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_6323 = transpose(perm = var_6322, x = var_6317)[name = string("transpose_28")];
+            tensor<fp16, [1, 8, 256]> x_303 = reshape(shape = var_6332, x = var_6323)[name = string("x_303")];
+            int32 var_6338 = const()[name = string("op_6338"), val = int32(-1)];
+            fp16 const_178_promoted_to_fp16 = const()[name = string("const_178_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_6344_cast_fp16 = mul(x = x_303, y = const_178_promoted_to_fp16)[name = string("op_6344_cast_fp16")];
+            bool input_297_interleave_0 = const()[name = string("input_297_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_297_cast_fp16 = concat(axis = var_6338, interleave = input_297_interleave_0, values = (x_303, var_6344_cast_fp16))[name = string("input_297_cast_fp16")];
+            tensor<int32, [1]> normed_285_axes_0 = const()[name = string("normed_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_285_cast_fp16 = layer_norm(axes = normed_285_axes_0, epsilon = var_6336_to_fp16, x = input_297_cast_fp16)[name = string("normed_285_cast_fp16")];
+            tensor<int32, [2]> var_6349_split_sizes_0 = const()[name = string("op_6349_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_6349_axis_0 = const()[name = string("op_6349_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_6349_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_6349_cast_fp16_1 = split(axis = var_6349_axis_0, split_sizes = var_6349_split_sizes_0, x = normed_285_cast_fp16)[name = string("op_6349_cast_fp16")];
+            tensor<fp16, [1, 8, 256]> var_6352_cast_fp16 = mul(x = var_6349_cast_fp16_0, y = const_57_to_fp16)[name = string("op_6352_cast_fp16")];
+            tensor<int32, [4]> var_6358 = const()[name = string("op_6358"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_83 = reshape(shape = var_6358, x = var_6352_cast_fp16)[name = string("q_83")];
+            tensor<fp16, [1, 8, 1, 256]> var_6360 = mul(x = q_83, y = cos_1)[name = string("op_6360")];
+            tensor<int32, [2]> var_6361_split_sizes_0 = const()[name = string("op_6361_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_6361_axis_0 = const()[name = string("op_6361_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_6361_0, tensor<fp16, [1, 8, 1, 128]> var_6361_1 = split(axis = var_6361_axis_0, split_sizes = var_6361_split_sizes_0, x = q_83)[name = string("op_6361")];
+            fp16 const_180_promoted = const()[name = string("const_180_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_6363 = mul(x = var_6361_1, y = const_180_promoted)[name = string("op_6363")];
+            int32 var_6365 = const()[name = string("op_6365"), val = int32(-1)];
+            bool var_6366_interleave_0 = const()[name = string("op_6366_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_6366 = concat(axis = var_6365, interleave = var_6366_interleave_0, values = (var_6363, var_6361_0))[name = string("op_6366")];
+            tensor<fp16, [1, 8, 1, 256]> var_6367 = mul(x = var_6366, y = sin_1)[name = string("op_6367")];
+            tensor<fp16, [1, 8, 1, 256]> q_87 = add(x = var_6360, y = var_6367)[name = string("q_87")];
+            string var_6380_pad_type_0 = const()[name = string("op_6380_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6380_strides_0 = const()[name = string("op_6380_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6380_pad_0 = const()[name = string("op_6380_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6380_dilations_0 = const()[name = string("op_6380_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6380_groups_0 = const()[name = string("op_6380_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_6380 = conv(dilations = var_6380_dilations_0, groups = var_6380_groups_0, pad = var_6380_pad_0, pad_type = var_6380_pad_type_0, strides = var_6380_strides_0, weight = layers_10_self_attn_k_proj_weight_palettized, x = var_6295_cast_fp16)[name = string("op_6380")];
+            tensor<int32, [4]> var_6385 = const()[name = string("op_6385"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_6386 = reshape(shape = var_6385, x = var_6380)[name = string("op_6386")];
+            tensor<int32, [4]> var_6391 = const()[name = string("op_6391"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_6408_pad_type_0 = const()[name = string("op_6408_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6408_strides_0 = const()[name = string("op_6408_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6408_pad_0 = const()[name = string("op_6408_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6408_dilations_0 = const()[name = string("op_6408_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6408_groups_0 = const()[name = string("op_6408_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_6408 = conv(dilations = var_6408_dilations_0, groups = var_6408_groups_0, pad = var_6408_pad_0, pad_type = var_6408_pad_type_0, strides = var_6408_strides_0, weight = layers_10_self_attn_v_proj_weight_palettized, x = var_6295_cast_fp16)[name = string("op_6408")];
+            tensor<int32, [4]> var_6413 = const()[name = string("op_6413"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_6414 = reshape(shape = var_6413, x = var_6408)[name = string("op_6414")];
+            tensor<int32, [4]> var_6419 = const()[name = string("op_6419"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_6429 = const()[name = string("op_6429"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_6392 = transpose(perm = var_6391, x = var_6386)[name = string("transpose_27")];
+            tensor<fp16, [1, 1, 256]> x_307 = reshape(shape = var_6429, x = var_6392)[name = string("x_307")];
+            int32 var_6435 = const()[name = string("op_6435"), val = int32(-1)];
+            fp16 const_181_promoted_to_fp16 = const()[name = string("const_181_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_6441_cast_fp16 = mul(x = x_307, y = const_181_promoted_to_fp16)[name = string("op_6441_cast_fp16")];
+            bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_299_cast_fp16 = concat(axis = var_6435, interleave = input_299_interleave_0, values = (x_307, var_6441_cast_fp16))[name = string("input_299_cast_fp16")];
+            tensor<int32, [1]> normed_289_axes_0 = const()[name = string("normed_289_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6433_to_fp16 = const()[name = string("op_6433_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_289_cast_fp16 = layer_norm(axes = normed_289_axes_0, epsilon = var_6433_to_fp16, x = input_299_cast_fp16)[name = string("normed_289_cast_fp16")];
+            tensor<int32, [2]> var_6446_split_sizes_0 = const()[name = string("op_6446_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_6446_axis_0 = const()[name = string("op_6446_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_6446_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_6446_cast_fp16_1 = split(axis = var_6446_axis_0, split_sizes = var_6446_split_sizes_0, x = normed_289_cast_fp16)[name = string("op_6446_cast_fp16")];
+            tensor<fp16, [256]> const_182_to_fp16 = const()[name = string("const_182_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221809600)))];
+            tensor<fp16, [1, 1, 256]> var_6449_cast_fp16 = mul(x = var_6446_cast_fp16_0, y = const_182_to_fp16)[name = string("op_6449_cast_fp16")];
+            tensor<int32, [4]> var_6455 = const()[name = string("op_6455"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_85 = reshape(shape = var_6455, x = var_6449_cast_fp16)[name = string("q_85")];
+            fp16 var_6462_promoted_to_fp16 = const()[name = string("op_6462_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_6420 = transpose(perm = var_6419, x = var_6414)[name = string("transpose_26")];
+            tensor<fp16, [1, 1, 1, 256]> var_6463_cast_fp16 = pow(x = var_6420, y = var_6462_promoted_to_fp16)[name = string("op_6463_cast_fp16")];
+            tensor<int32, [1]> var_6468_axes_0 = const()[name = string("op_6468_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_6468_keep_dims_0 = const()[name = string("op_6468_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_6468_cast_fp16 = reduce_mean(axes = var_6468_axes_0, keep_dims = var_6468_keep_dims_0, x = var_6463_cast_fp16)[name = string("op_6468_cast_fp16")];
+            fp16 var_6470_to_fp16 = const()[name = string("op_6470_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_21_cast_fp16 = add(x = var_6468_cast_fp16, y = var_6470_to_fp16)[name = string("mean_sq_21_cast_fp16")];
+            fp16 var_6477_to_fp16 = const()[name = string("op_6477_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_6478_cast_fp16 = pow(x = mean_sq_21_cast_fp16, y = var_6477_to_fp16)[name = string("op_6478_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_6479_cast_fp16 = mul(x = var_6420, y = var_6478_cast_fp16)[name = string("op_6479_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_6485 = mul(x = q_85, y = cos_1)[name = string("op_6485")];
+            tensor<int32, [2]> var_6486_split_sizes_0 = const()[name = string("op_6486_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_6486_axis_0 = const()[name = string("op_6486_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_6486_0, tensor<fp16, [1, 1, 1, 128]> var_6486_1 = split(axis = var_6486_axis_0, split_sizes = var_6486_split_sizes_0, x = q_85)[name = string("op_6486")];
+            fp16 const_183_promoted = const()[name = string("const_183_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_6488 = mul(x = var_6486_1, y = const_183_promoted)[name = string("op_6488")];
+            int32 var_6490 = const()[name = string("op_6490"), val = int32(-1)];
+            bool var_6491_interleave_0 = const()[name = string("op_6491_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_6491 = concat(axis = var_6490, interleave = var_6491_interleave_0, values = (var_6488, var_6486_0))[name = string("op_6491")];
+            tensor<fp16, [1, 1, 1, 256]> var_6492 = mul(x = var_6491, y = sin_1)[name = string("op_6492")];
+            tensor<fp16, [1, 1, 1, 256]> input_301 = add(x = var_6485, y = var_6492)[name = string("input_301")];
+            tensor<int32, [4]> var_6497_begin_0 = const()[name = string("op_6497_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> var_6497_end_0 = const()[name = string("op_6497_end_0"), val = tensor<int32, [4]>([11, 1, 512, 512])];
+            tensor<bool, [4]> var_6497_end_mask_0 = const()[name = string("op_6497_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_6497_squeeze_mask_0 = const()[name = string("op_6497_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_6497_cast_fp16 = slice_by_index(begin = var_6497_begin_0, end = var_6497_end_0, end_mask = var_6497_end_mask_0, squeeze_mask = var_6497_squeeze_mask_0, x = coreml_update_state_43)[name = string("op_6497_cast_fp16")];
+            tensor<int32, [1]> K_c_21_axes_0 = const()[name = string("K_c_21_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_21_cast_fp16 = expand_dims(axes = K_c_21_axes_0, x = var_6497_cast_fp16)[name = string("K_c_21_cast_fp16")];
+            tensor<int32, [4]> var_6502_begin_0 = const()[name = string("op_6502_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> var_6502_end_0 = const()[name = string("op_6502_end_0"), val = tensor<int32, [4]>([23, 1, 512, 512])];
+            tensor<bool, [4]> var_6502_end_mask_0 = const()[name = string("op_6502_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_6502_squeeze_mask_0 = const()[name = string("op_6502_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_6502_cast_fp16 = slice_by_index(begin = var_6502_begin_0, end = var_6502_end_0, end_mask = var_6502_end_mask_0, squeeze_mask = var_6502_squeeze_mask_0, x = coreml_update_state_43)[name = string("op_6502_cast_fp16")];
+            tensor<int32, [1]> V_c_21_axes_0 = const()[name = string("V_c_21_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_21_cast_fp16 = expand_dims(axes = V_c_21_axes_0, x = var_6502_cast_fp16)[name = string("V_c_21_cast_fp16")];
+            tensor<int32, [8]> kp_17_pad_0 = const()[name = string("kp_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_17_mode_0 = const()[name = string("kp_17_mode_0"), val = string("constant")];
+            fp16 const_184_to_fp16 = const()[name = string("const_184_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_17_cast_fp16 = pad(constant_val = const_184_to_fp16, mode = kp_17_mode_0, pad = kp_17_pad_0, x = input_301)[name = string("kp_17_cast_fp16")];
+            tensor<int32, [8]> vp_17_pad_0 = const()[name = string("vp_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_17_mode_0 = const()[name = string("vp_17_mode_0"), val = string("constant")];
+            fp16 const_185_to_fp16 = const()[name = string("const_185_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_17_cast_fp16 = pad(constant_val = const_185_to_fp16, mode = vp_17_mode_0, pad = vp_17_pad_0, x = var_6479_cast_fp16)[name = string("vp_17_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_6520_cast_fp16 = mul(x = K_c_21_cast_fp16, y = var_1005_cast_fp16)[name = string("op_6520_cast_fp16")];
+            tensor<int32, [4]> var_6521_reps_0 = const()[name = string("op_6521_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_6521_cast_fp16 = tile(reps = var_6521_reps_0, x = kp_17_cast_fp16)[name = string("op_6521_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_6522_cast_fp16 = mul(x = var_6521_cast_fp16, y = update_mask)[name = string("op_6522_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_21_cast_fp16 = add(x = var_6520_cast_fp16, y = var_6522_cast_fp16)[name = string("K_n_21_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_6528_cast_fp16 = mul(x = V_c_21_cast_fp16, y = var_1005_cast_fp16)[name = string("op_6528_cast_fp16")];
+            tensor<int32, [4]> var_6529_reps_0 = const()[name = string("op_6529_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_6529_cast_fp16 = tile(reps = var_6529_reps_0, x = vp_17_cast_fp16)[name = string("op_6529_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_6530_cast_fp16 = mul(x = var_6529_cast_fp16, y = update_mask)[name = string("op_6530_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_21_cast_fp16 = add(x = var_6528_cast_fp16, y = var_6530_cast_fp16)[name = string("V_n_21_cast_fp16")];
+            tensor<int32, [1]> var_6534_axes_0 = const()[name = string("op_6534_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_6534_cast_fp16 = squeeze(axes = var_6534_axes_0, x = K_n_21_cast_fp16)[name = string("op_6534_cast_fp16")];
+            tensor<int32, [4]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_21_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_21_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_21_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_80, begin_mask = kv_cache_0_internal_tensor_assign_21_begin_mask_0, end = concat_81, end_mask = kv_cache_0_internal_tensor_assign_21_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_21_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_21_stride_0, update = var_6534_cast_fp16, x = coreml_update_state_43)[name = string("kv_cache_0_internal_tensor_assign_21_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_21_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_44_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_44 = read_state(input = kv_cache_0)[name = string("coreml_update_state_44")];
+            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_6541_cast_fp16 = squeeze(axes = var_6541_axes_0, x = V_n_21_cast_fp16)[name = string("op_6541_cast_fp16")];
+            tensor<int32, [4]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_22_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_22_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_22_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_22_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_82, begin_mask = kv_cache_0_internal_tensor_assign_22_begin_mask_0, end = concat_83, end_mask = kv_cache_0_internal_tensor_assign_22_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_22_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_22_stride_0, update = var_6541_cast_fp16, x = coreml_update_state_44)[name = string("kv_cache_0_internal_tensor_assign_22_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_22_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_45_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_45 = read_state(input = kv_cache_0)[name = string("coreml_update_state_45")];
+            tensor<int32, [4]> var_6551_begin_0 = const()[name = string("op_6551_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_6551_end_0 = const()[name = string("op_6551_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_6551_end_mask_0 = const()[name = string("op_6551_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_6551_cast_fp16 = slice_by_index(begin = var_6551_begin_0, end = var_6551_end_0, end_mask = var_6551_end_mask_0, x = K_n_21_cast_fp16)[name = string("op_6551_cast_fp16")];
+            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = var_6551_cast_fp16)[name = string("transpose_25")];
+            tensor<fp16, [8, 1, 512, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
+            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_40_cast_fp16 = reshape(shape = concat_84, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
+            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_24")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_41_cast_fp16 = reshape(shape = concat_85, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
+            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_6560_begin_0 = const()[name = string("op_6560_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_6560_end_0 = const()[name = string("op_6560_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_6560_end_mask_0 = const()[name = string("op_6560_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_6560_cast_fp16 = slice_by_index(begin = var_6560_begin_0, end = var_6560_end_0, end_mask = var_6560_end_mask_0, x = V_n_21_cast_fp16)[name = string("op_6560_cast_fp16")];
+            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = var_6560_cast_fp16)[name = string("transpose_23")];
+            tensor<fp16, [8, 1, 512, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
+            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_42_cast_fp16 = reshape(shape = concat_86, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
+            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_22")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_43_cast_fp16 = reshape(shape = concat_87, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
+            tensor<int32, [4]> Ve_21_perm_0 = const()[name = string("Ve_21_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_6578_transpose_x_0 = const()[name = string("op_6578_transpose_x_0"), val = bool(false)];
+            bool var_6578_transpose_y_0 = const()[name = string("op_6578_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_21")];
+            tensor<fp16, [1, 8, 1, 512]> var_6578_cast_fp16 = matmul(transpose_x = var_6578_transpose_x_0, transpose_y = var_6578_transpose_y_0, x = q_87, y = transpose_58_cast_fp16)[name = string("op_6578_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_6585_cast_fp16 = add(x = var_6578_cast_fp16, y = causal_mask)[name = string("op_6585_cast_fp16")];
+            int32 var_6586 = const()[name = string("op_6586"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_6588_cast_fp16 = softmax(axis = var_6586, x = var_6585_cast_fp16)[name = string("op_6588_cast_fp16")];
+            bool var_6604_transpose_x_0 = const()[name = string("op_6604_transpose_x_0"), val = bool(false)];
+            bool var_6604_transpose_y_0 = const()[name = string("op_6604_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_21_cast_fp16 = transpose(perm = Ve_21_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_20")];
+            tensor<fp16, [1, 8, 1, 256]> var_6604_cast_fp16 = matmul(transpose_x = var_6604_transpose_x_0, transpose_y = var_6604_transpose_y_0, x = var_6588_cast_fp16, y = Ve_21_cast_fp16)[name = string("op_6604_cast_fp16")];
+            tensor<int32, [4]> var_6614 = const()[name = string("op_6614"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_6621 = const()[name = string("op_6621"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_6615 = transpose(perm = var_6614, x = var_6604_cast_fp16)[name = string("transpose_19")];
+            tensor<fp16, [1, 1, 2048]> var_6622 = reshape(shape = var_6621, x = var_6615)[name = string("op_6622")];
+            tensor<int32, [3]> var_6626 = const()[name = string("op_6626"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_10_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221810176))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223383104))))[name = string("squeeze_10_palettized")];
+            string var_6642_pad_type_0 = const()[name = string("op_6642_pad_type_0"), val = string("valid")];
+            int32 var_6642_groups_0 = const()[name = string("op_6642_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_6642_strides_0 = const()[name = string("op_6642_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_6642_pad_0 = const()[name = string("op_6642_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_6642_dilations_0 = const()[name = string("op_6642_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_6627 = transpose(perm = var_6626, x = var_6622)[name = string("transpose_18")];
+            tensor<fp16, [1, 1536, 1]> var_6642 = conv(dilations = var_6642_dilations_0, groups = var_6642_groups_0, pad = var_6642_pad_0, pad_type = var_6642_pad_type_0, strides = var_6642_strides_0, weight = squeeze_10_palettized, x = var_6627)[name = string("op_6642")];
+            tensor<int32, [3]> var_6646 = const()[name = string("op_6646"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_6652 = const()[name = string("op_6652"), val = int32(-1)];
+            fp16 const_186_promoted_to_fp16 = const()[name = string("const_186_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_313 = transpose(perm = var_6646, x = var_6642)[name = string("transpose_17")];
+            tensor<fp16, [1, 1, 1536]> var_6658_cast_fp16 = mul(x = x_313, y = const_186_promoted_to_fp16)[name = string("op_6658_cast_fp16")];
+            bool input_307_interleave_0 = const()[name = string("input_307_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_307_cast_fp16 = concat(axis = var_6652, interleave = input_307_interleave_0, values = (x_313, var_6658_cast_fp16))[name = string("input_307_cast_fp16")];
+            tensor<int32, [1]> normed_293_axes_0 = const()[name = string("normed_293_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6650_to_fp16 = const()[name = string("op_6650_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_293_cast_fp16 = layer_norm(axes = normed_293_axes_0, epsilon = var_6650_to_fp16, x = input_307_cast_fp16)[name = string("normed_293_cast_fp16")];
+            tensor<int32, [2]> var_6663_split_sizes_0 = const()[name = string("op_6663_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6663_axis_0 = const()[name = string("op_6663_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6663_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6663_cast_fp16_1 = split(axis = var_6663_axis_0, split_sizes = var_6663_split_sizes_0, x = normed_293_cast_fp16)[name = string("op_6663_cast_fp16")];
+            tensor<fp16, [1536]> const_187_to_fp16 = const()[name = string("const_187_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223384704)))];
+            tensor<fp16, [1, 1, 1536]> var_6666_cast_fp16 = mul(x = var_6663_cast_fp16_0, y = const_187_to_fp16)[name = string("op_6666_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_317_cast_fp16 = add(x = x_299_cast_fp16, y = var_6666_cast_fp16)[name = string("x_317_cast_fp16")];
+            int32 var_6673 = const()[name = string("op_6673"), val = int32(-1)];
+            fp16 const_188_promoted_to_fp16 = const()[name = string("const_188_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_6679_cast_fp16 = mul(x = x_317_cast_fp16, y = const_188_promoted_to_fp16)[name = string("op_6679_cast_fp16")];
+            bool input_309_interleave_0 = const()[name = string("input_309_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_309_cast_fp16 = concat(axis = var_6673, interleave = input_309_interleave_0, values = (x_317_cast_fp16, var_6679_cast_fp16))[name = string("input_309_cast_fp16")];
+            tensor<int32, [1]> normed_297_axes_0 = const()[name = string("normed_297_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6671_to_fp16 = const()[name = string("op_6671_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_297_cast_fp16 = layer_norm(axes = normed_297_axes_0, epsilon = var_6671_to_fp16, x = input_309_cast_fp16)[name = string("normed_297_cast_fp16")];
+            tensor<int32, [2]> var_6684_split_sizes_0 = const()[name = string("op_6684_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6684_axis_0 = const()[name = string("op_6684_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6684_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6684_cast_fp16_1 = split(axis = var_6684_axis_0, split_sizes = var_6684_split_sizes_0, x = normed_297_cast_fp16)[name = string("op_6684_cast_fp16")];
+            tensor<fp16, [1536]> const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223387840)))];
+            tensor<fp16, [1, 1, 1536]> var_6687_cast_fp16 = mul(x = var_6684_cast_fp16_0, y = const_189_to_fp16)[name = string("op_6687_cast_fp16")];
+            tensor<int32, [3]> var_6700 = const()[name = string("op_6700"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_311_axes_0 = const()[name = string("input_311_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_6701 = transpose(perm = var_6700, x = var_6687_cast_fp16)[name = string("transpose_16")];
+            tensor<fp16, [1, 1536, 1, 1]> input_311 = expand_dims(axes = input_311_axes_0, x = var_6701)[name = string("input_311")];
+            string var_6714_pad_type_0 = const()[name = string("op_6714_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6714_strides_0 = const()[name = string("op_6714_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6714_pad_0 = const()[name = string("op_6714_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6714_dilations_0 = const()[name = string("op_6714_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6714_groups_0 = const()[name = string("op_6714_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_6714 = conv(dilations = var_6714_dilations_0, groups = var_6714_groups_0, pad = var_6714_pad_0, pad_type = var_6714_pad_type_0, strides = var_6714_strides_0, weight = layers_10_mlp_gate_proj_weight_palettized, x = input_311)[name = string("op_6714")];
+            string var_6716_mode_0 = const()[name = string("op_6716_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_6716 = gelu(mode = var_6716_mode_0, x = var_6714)[name = string("op_6716")];
+            string var_6727_pad_type_0 = const()[name = string("op_6727_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6727_strides_0 = const()[name = string("op_6727_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6727_pad_0 = const()[name = string("op_6727_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6727_dilations_0 = const()[name = string("op_6727_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6727_groups_0 = const()[name = string("op_6727_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_6727 = conv(dilations = var_6727_dilations_0, groups = var_6727_groups_0, pad = var_6727_pad_0, pad_type = var_6727_pad_type_0, strides = var_6727_strides_0, weight = layers_10_mlp_up_proj_weight_palettized, x = input_311)[name = string("op_6727")];
+            tensor<fp16, [1, 6144, 1, 1]> input_313 = mul(x = var_6716, y = var_6727)[name = string("input_313")];
+            string var_6739_pad_type_0 = const()[name = string("op_6739_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6739_strides_0 = const()[name = string("op_6739_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6739_pad_0 = const()[name = string("op_6739_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6739_dilations_0 = const()[name = string("op_6739_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6739_groups_0 = const()[name = string("op_6739_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_6739 = conv(dilations = var_6739_dilations_0, groups = var_6739_groups_0, pad = var_6739_pad_0, pad_type = var_6739_pad_type_0, strides = var_6739_strides_0, weight = layers_10_mlp_down_proj_weight_palettized, x = input_313)[name = string("op_6739")];
+            tensor<int32, [1]> var_6741_axes_0 = const()[name = string("op_6741_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_6741 = squeeze(axes = var_6741_axes_0, x = var_6739)[name = string("op_6741")];
+            tensor<int32, [3]> var_6745 = const()[name = string("op_6745"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_6751 = const()[name = string("op_6751"), val = int32(-1)];
+            fp16 const_190_promoted_to_fp16 = const()[name = string("const_190_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_321 = transpose(perm = var_6745, x = var_6741)[name = string("transpose_15")];
+            tensor<fp16, [1, 1, 1536]> var_6757_cast_fp16 = mul(x = x_321, y = const_190_promoted_to_fp16)[name = string("op_6757_cast_fp16")];
+            bool input_315_interleave_0 = const()[name = string("input_315_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_315_cast_fp16 = concat(axis = var_6751, interleave = input_315_interleave_0, values = (x_321, var_6757_cast_fp16))[name = string("input_315_cast_fp16")];
+            tensor<int32, [1]> normed_301_axes_0 = const()[name = string("normed_301_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6749_to_fp16 = const()[name = string("op_6749_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_301_cast_fp16 = layer_norm(axes = normed_301_axes_0, epsilon = var_6749_to_fp16, x = input_315_cast_fp16)[name = string("normed_301_cast_fp16")];
+            tensor<int32, [2]> var_6762_split_sizes_0 = const()[name = string("op_6762_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6762_axis_0 = const()[name = string("op_6762_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6762_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6762_cast_fp16_1 = split(axis = var_6762_axis_0, split_sizes = var_6762_split_sizes_0, x = normed_301_cast_fp16)[name = string("op_6762_cast_fp16")];
+            tensor<fp16, [1536]> const_191_to_fp16 = const()[name = string("const_191_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223390976)))];
+            tensor<fp16, [1, 1, 1536]> var_6765_cast_fp16 = mul(x = var_6762_cast_fp16_0, y = const_191_to_fp16)[name = string("op_6765_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_153_cast_fp16 = add(x = x_317_cast_fp16, y = var_6765_cast_fp16)[name = string("hidden_states_153_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_6776 = linear(bias = linear_0_bias_0, weight = layers_10_per_layer_input_gate_weight_palettized, x = hidden_states_153_cast_fp16)[name = string("linear_20")];
+            string gated_21_mode_0 = const()[name = string("gated_21_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated_21 = gelu(mode = gated_21_mode_0, x = var_6776)[name = string("gated_21")];
+            tensor<int32, [3]> var_6793_begin_0 = const()[name = string("op_6793_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
+            tensor<int32, [3]> var_6793_end_0 = const()[name = string("op_6793_end_0"), val = tensor<int32, [3]>([1, 1, 2816])];
+            tensor<bool, [3]> var_6793_end_mask_0 = const()[name = string("op_6793_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_6793_cast_fp16 = slice_by_index(begin = var_6793_begin_0, end = var_6793_end_0, end_mask = var_6793_end_mask_0, x = per_layer_combined)[name = string("op_6793_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_319_cast_fp16 = mul(x = gated_21, y = var_6793_cast_fp16)[name = string("input_319_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_10_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223394112))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223590784))))[name = string("layers_10_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_319_cast_fp16)[name = string("linear_21_cast_fp16")];
+            int32 var_6802 = const()[name = string("op_6802"), val = int32(-1)];
+            fp16 const_192_promoted_to_fp16 = const()[name = string("const_192_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_6808_cast_fp16 = mul(x = linear_21_cast_fp16, y = const_192_promoted_to_fp16)[name = string("op_6808_cast_fp16")];
+            bool input_321_interleave_0 = const()[name = string("input_321_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_321_cast_fp16 = concat(axis = var_6802, interleave = input_321_interleave_0, values = (linear_21_cast_fp16, var_6808_cast_fp16))[name = string("input_321_cast_fp16")];
+            tensor<int32, [1]> normed_305_axes_0 = const()[name = string("normed_305_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6800_to_fp16 = const()[name = string("op_6800_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_305_cast_fp16 = layer_norm(axes = normed_305_axes_0, epsilon = var_6800_to_fp16, x = input_321_cast_fp16)[name = string("normed_305_cast_fp16")];
+            tensor<int32, [2]> var_6813_split_sizes_0 = const()[name = string("op_6813_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6813_axis_0 = const()[name = string("op_6813_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6813_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6813_cast_fp16_1 = split(axis = var_6813_axis_0, split_sizes = var_6813_split_sizes_0, x = normed_305_cast_fp16)[name = string("op_6813_cast_fp16")];
+            tensor<fp16, [1536]> const_193_to_fp16 = const()[name = string("const_193_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223592384)))];
+            tensor<fp16, [1, 1, 1536]> var_6816_cast_fp16 = mul(x = var_6813_cast_fp16_0, y = const_193_to_fp16)[name = string("op_6816_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_157_cast_fp16 = add(x = hidden_states_153_cast_fp16, y = var_6816_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
+            tensor<fp16, [1]> layers_10_layer_scalar_to_fp16 = const()[name = string("layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.c6p-2])];
+            tensor<fp16, [1, 1, 1536]> x_329_cast_fp16 = mul(x = hidden_states_157_cast_fp16, y = layers_10_layer_scalar_to_fp16)[name = string("x_329_cast_fp16")];
+            int32 var_6824 = const()[name = string("op_6824"), 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_6830_cast_fp16 = mul(x = x_329_cast_fp16, y = const_194_promoted_to_fp16)[name = string("op_6830_cast_fp16")];
+            bool input_323_interleave_0 = const()[name = string("input_323_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_323_cast_fp16 = concat(axis = var_6824, interleave = input_323_interleave_0, values = (x_329_cast_fp16, var_6830_cast_fp16))[name = string("input_323_cast_fp16")];
+            tensor<int32, [1]> normed_309_axes_0 = const()[name = string("normed_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6822_to_fp16 = const()[name = string("op_6822_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_309_cast_fp16 = layer_norm(axes = normed_309_axes_0, epsilon = var_6822_to_fp16, x = input_323_cast_fp16)[name = string("normed_309_cast_fp16")];
+            tensor<int32, [2]> var_6835_split_sizes_0 = const()[name = string("op_6835_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_6835_axis_0 = const()[name = string("op_6835_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_6835_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_6835_cast_fp16_1 = split(axis = var_6835_axis_0, split_sizes = var_6835_split_sizes_0, x = normed_309_cast_fp16)[name = string("op_6835_cast_fp16")];
+            tensor<fp16, [1536]> const_195_to_fp16 = const()[name = string("const_195_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223595520)))];
+            tensor<fp16, [1, 1, 1536]> var_6838_cast_fp16 = mul(x = var_6835_cast_fp16_0, y = const_195_to_fp16)[name = string("op_6838_cast_fp16")];
+            tensor<int32, [3]> var_6846 = const()[name = string("op_6846"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_6849_axes_0 = const()[name = string("op_6849_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_6847_cast_fp16 = transpose(perm = var_6846, x = var_6838_cast_fp16)[name = string("transpose_14")];
+            tensor<fp16, [1, 1536, 1, 1]> var_6849_cast_fp16 = expand_dims(axes = var_6849_axes_0, x = var_6847_cast_fp16)[name = string("op_6849_cast_fp16")];
+            string var_6865_pad_type_0 = const()[name = string("op_6865_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6865_strides_0 = const()[name = string("op_6865_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6865_pad_0 = const()[name = string("op_6865_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6865_dilations_0 = const()[name = string("op_6865_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6865_groups_0 = const()[name = string("op_6865_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> var_6865 = conv(dilations = var_6865_dilations_0, groups = var_6865_groups_0, pad = var_6865_pad_0, pad_type = var_6865_pad_type_0, strides = var_6865_strides_0, weight = layers_11_self_attn_q_proj_weight_palettized, x = var_6849_cast_fp16)[name = string("op_6865")];
+            tensor<int32, [4]> var_6870 = const()[name = string("op_6870"), val = tensor<int32, [4]>([1, 8, 256, 1])];
+            tensor<fp16, [1, 8, 256, 1]> var_6871 = reshape(shape = var_6870, x = var_6865)[name = string("op_6871")];
+            tensor<int32, [4]> var_6876 = const()[name = string("op_6876"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_6886 = const()[name = string("op_6886"), val = tensor<int32, [3]>([1, 8, 256])];
+            tensor<fp16, [1, 8, 1, 256]> var_6877 = transpose(perm = var_6876, x = var_6871)[name = string("transpose_13")];
+            tensor<fp16, [1, 8, 256]> x_333 = reshape(shape = var_6886, x = var_6877)[name = string("x_333")];
+            int32 var_6892 = const()[name = string("op_6892"), val = int32(-1)];
+            fp16 const_196_promoted_to_fp16 = const()[name = string("const_196_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 256]> var_6898_cast_fp16 = mul(x = x_333, y = const_196_promoted_to_fp16)[name = string("op_6898_cast_fp16")];
+            bool input_327_interleave_0 = const()[name = string("input_327_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512]> input_327_cast_fp16 = concat(axis = var_6892, interleave = input_327_interleave_0, values = (x_333, var_6898_cast_fp16))[name = string("input_327_cast_fp16")];
+            tensor<int32, [1]> normed_313_axes_0 = const()[name = string("normed_313_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6890_to_fp16 = const()[name = string("op_6890_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512]> normed_313_cast_fp16 = layer_norm(axes = normed_313_axes_0, epsilon = var_6890_to_fp16, x = input_327_cast_fp16)[name = string("normed_313_cast_fp16")];
+            tensor<int32, [2]> var_6903_split_sizes_0 = const()[name = string("op_6903_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_6903_axis_0 = const()[name = string("op_6903_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 256]> var_6903_cast_fp16_0, tensor<fp16, [1, 8, 256]> var_6903_cast_fp16_1 = split(axis = var_6903_axis_0, split_sizes = var_6903_split_sizes_0, x = normed_313_cast_fp16)[name = string("op_6903_cast_fp16")];
+            tensor<int32, [4]> var_6912 = const()[name = string("op_6912"), val = tensor<int32, [4]>([1, 8, 1, 256])];
+            tensor<fp16, [1, 8, 1, 256]> q_91 = reshape(shape = var_6912, x = var_6903_cast_fp16_0)[name = string("q_91")];
+            tensor<fp16, [1, 8, 1, 256]> var_6914 = mul(x = q_91, y = cos_1)[name = string("op_6914")];
+            tensor<int32, [2]> var_6915_split_sizes_0 = const()[name = string("op_6915_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_6915_axis_0 = const()[name = string("op_6915_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 128]> var_6915_0, tensor<fp16, [1, 8, 1, 128]> var_6915_1 = split(axis = var_6915_axis_0, split_sizes = var_6915_split_sizes_0, x = q_91)[name = string("op_6915")];
+            fp16 const_198_promoted = const()[name = string("const_198_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 8, 1, 128]> var_6917 = mul(x = var_6915_1, y = const_198_promoted)[name = string("op_6917")];
+            int32 var_6919 = const()[name = string("op_6919"), val = int32(-1)];
+            bool var_6920_interleave_0 = const()[name = string("op_6920_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1, 256]> var_6920 = concat(axis = var_6919, interleave = var_6920_interleave_0, values = (var_6917, var_6915_0))[name = string("op_6920")];
+            tensor<fp16, [1, 8, 1, 256]> var_6921 = mul(x = var_6920, y = sin_1)[name = string("op_6921")];
+            tensor<fp16, [1, 8, 1, 256]> q = add(x = var_6914, y = var_6921)[name = string("q")];
+            string var_6934_pad_type_0 = const()[name = string("op_6934_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6934_strides_0 = const()[name = string("op_6934_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6934_pad_0 = const()[name = string("op_6934_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6934_dilations_0 = const()[name = string("op_6934_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6934_groups_0 = const()[name = string("op_6934_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_6934 = conv(dilations = var_6934_dilations_0, groups = var_6934_groups_0, pad = var_6934_pad_0, pad_type = var_6934_pad_type_0, strides = var_6934_strides_0, weight = layers_11_self_attn_k_proj_weight_palettized, x = var_6849_cast_fp16)[name = string("op_6934")];
+            tensor<int32, [4]> var_6939 = const()[name = string("op_6939"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_6940 = reshape(shape = var_6939, x = var_6934)[name = string("op_6940")];
+            tensor<int32, [4]> var_6945 = const()[name = string("op_6945"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            string var_6962_pad_type_0 = const()[name = string("op_6962_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_6962_strides_0 = const()[name = string("op_6962_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6962_pad_0 = const()[name = string("op_6962_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6962_dilations_0 = const()[name = string("op_6962_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_6962_groups_0 = const()[name = string("op_6962_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 256, 1, 1]> var_6962 = conv(dilations = var_6962_dilations_0, groups = var_6962_groups_0, pad = var_6962_pad_0, pad_type = var_6962_pad_type_0, strides = var_6962_strides_0, weight = layers_11_self_attn_v_proj_weight_palettized, x = var_6849_cast_fp16)[name = string("op_6962")];
+            tensor<int32, [4]> var_6967 = const()[name = string("op_6967"), val = tensor<int32, [4]>([1, 1, 256, 1])];
+            tensor<fp16, [1, 1, 256, 1]> var_6968 = reshape(shape = var_6967, x = var_6962)[name = string("op_6968")];
+            tensor<int32, [4]> var_6973 = const()[name = string("op_6973"), val = tensor<int32, [4]>([0, 1, 3, 2])];
+            tensor<int32, [3]> var_6983 = const()[name = string("op_6983"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_6946 = transpose(perm = var_6945, x = var_6940)[name = string("transpose_12")];
+            tensor<fp16, [1, 1, 256]> x_337 = reshape(shape = var_6983, x = var_6946)[name = string("x_337")];
+            int32 var_6989 = const()[name = string("op_6989"), val = int32(-1)];
+            fp16 const_199_promoted_to_fp16 = const()[name = string("const_199_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 256]> var_6995_cast_fp16 = mul(x = x_337, y = const_199_promoted_to_fp16)[name = string("op_6995_cast_fp16")];
+            bool input_329_interleave_0 = const()[name = string("input_329_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 512]> input_329_cast_fp16 = concat(axis = var_6989, interleave = input_329_interleave_0, values = (x_337, var_6995_cast_fp16))[name = string("input_329_cast_fp16")];
+            tensor<int32, [1]> normed_317_axes_0 = const()[name = string("normed_317_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_6987_to_fp16 = const()[name = string("op_6987_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 512]> normed_317_cast_fp16 = layer_norm(axes = normed_317_axes_0, epsilon = var_6987_to_fp16, x = input_329_cast_fp16)[name = string("normed_317_cast_fp16")];
+            tensor<int32, [2]> var_7000_split_sizes_0 = const()[name = string("op_7000_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            int32 var_7000_axis_0 = const()[name = string("op_7000_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 256]> var_7000_cast_fp16_0, tensor<fp16, [1, 1, 256]> var_7000_cast_fp16_1 = split(axis = var_7000_axis_0, split_sizes = var_7000_split_sizes_0, x = normed_317_cast_fp16)[name = string("op_7000_cast_fp16")];
+            tensor<fp16, [256]> const_200_to_fp16 = const()[name = string("const_200_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223598656)))];
+            tensor<fp16, [1, 1, 256]> var_7003_cast_fp16 = mul(x = var_7000_cast_fp16_0, y = const_200_to_fp16)[name = string("op_7003_cast_fp16")];
+            tensor<int32, [4]> var_7009 = const()[name = string("op_7009"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> q_93 = reshape(shape = var_7009, x = var_7003_cast_fp16)[name = string("q_93")];
+            fp16 var_7016_promoted_to_fp16 = const()[name = string("op_7016_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 1, 1, 256]> var_6974 = transpose(perm = var_6973, x = var_6968)[name = string("transpose_11")];
+            tensor<fp16, [1, 1, 1, 256]> var_7017_cast_fp16 = pow(x = var_6974, y = var_7016_promoted_to_fp16)[name = string("op_7017_cast_fp16")];
+            tensor<int32, [1]> var_7022_axes_0 = const()[name = string("op_7022_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_7022_keep_dims_0 = const()[name = string("op_7022_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> var_7022_cast_fp16 = reduce_mean(axes = var_7022_axes_0, keep_dims = var_7022_keep_dims_0, x = var_7017_cast_fp16)[name = string("op_7022_cast_fp16")];
+            fp16 var_7024_to_fp16 = const()[name = string("op_7024_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 1, 1]> mean_sq_cast_fp16 = add(x = var_7022_cast_fp16, y = var_7024_to_fp16)[name = string("mean_sq_cast_fp16")];
+            fp16 var_7031_to_fp16 = const()[name = string("op_7031_to_fp16"), val = fp16(-0x1p-1)];
+            tensor<fp16, [1, 1, 1, 1]> var_7032_cast_fp16 = pow(x = mean_sq_cast_fp16, y = var_7031_to_fp16)[name = string("op_7032_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_7033_cast_fp16 = mul(x = var_6974, y = var_7032_cast_fp16)[name = string("op_7033_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_7039 = mul(x = q_93, y = cos_1)[name = string("op_7039")];
+            tensor<int32, [2]> var_7040_split_sizes_0 = const()[name = string("op_7040_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
+            int32 var_7040_axis_0 = const()[name = string("op_7040_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 128]> var_7040_0, tensor<fp16, [1, 1, 1, 128]> var_7040_1 = split(axis = var_7040_axis_0, split_sizes = var_7040_split_sizes_0, x = q_93)[name = string("op_7040")];
+            fp16 const_201_promoted = const()[name = string("const_201_promoted"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1, 128]> var_7042 = mul(x = var_7040_1, y = const_201_promoted)[name = string("op_7042")];
+            int32 var_7044 = const()[name = string("op_7044"), val = int32(-1)];
+            bool var_7045_interleave_0 = const()[name = string("op_7045_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1, 256]> var_7045 = concat(axis = var_7044, interleave = var_7045_interleave_0, values = (var_7042, var_7040_0))[name = string("op_7045")];
+            tensor<fp16, [1, 1, 1, 256]> var_7046 = mul(x = var_7045, y = sin_1)[name = string("op_7046")];
+            tensor<fp16, [1, 1, 1, 256]> input_331 = add(x = var_7039, y = var_7046)[name = string("input_331")];
+            tensor<int32, [4]> var_7051_begin_0 = const()[name = string("op_7051_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> var_7051_end_0 = const()[name = string("op_7051_end_0"), val = tensor<int32, [4]>([12, 1, 512, 512])];
+            tensor<bool, [4]> var_7051_end_mask_0 = const()[name = string("op_7051_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_7051_squeeze_mask_0 = const()[name = string("op_7051_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_7051_cast_fp16 = slice_by_index(begin = var_7051_begin_0, end = var_7051_end_0, end_mask = var_7051_end_mask_0, squeeze_mask = var_7051_squeeze_mask_0, x = coreml_update_state_45)[name = string("op_7051_cast_fp16")];
+            tensor<int32, [1]> K_c_axes_0 = const()[name = string("K_c_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> K_c_cast_fp16 = expand_dims(axes = K_c_axes_0, x = var_7051_cast_fp16)[name = string("K_c_cast_fp16")];
+            tensor<int32, [4]> var_7056_begin_0 = const()[name = string("op_7056_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> var_7056_end_0 = const()[name = string("op_7056_end_0"), val = tensor<int32, [4]>([24, 1, 512, 512])];
+            tensor<bool, [4]> var_7056_end_mask_0 = const()[name = string("op_7056_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> var_7056_squeeze_mask_0 = const()[name = string("op_7056_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 512]> var_7056_cast_fp16 = slice_by_index(begin = var_7056_begin_0, end = var_7056_end_0, end_mask = var_7056_end_mask_0, squeeze_mask = var_7056_squeeze_mask_0, x = coreml_update_state_45)[name = string("op_7056_cast_fp16")];
+            tensor<int32, [1]> V_c_axes_0 = const()[name = string("V_c_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 512, 512]> V_c_cast_fp16 = expand_dims(axes = V_c_axes_0, x = var_7056_cast_fp16)[name = string("V_c_cast_fp16")];
+            tensor<int32, [8]> kp_pad_0 = const()[name = string("kp_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string kp_mode_0 = const()[name = string("kp_mode_0"), val = string("constant")];
+            fp16 const_202_to_fp16 = const()[name = string("const_202_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> kp_cast_fp16 = pad(constant_val = const_202_to_fp16, mode = kp_mode_0, pad = kp_pad_0, x = input_331)[name = string("kp_cast_fp16")];
+            tensor<int32, [8]> vp_pad_0 = const()[name = string("vp_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 256])];
+            string vp_mode_0 = const()[name = string("vp_mode_0"), val = string("constant")];
+            fp16 const_203_to_fp16 = const()[name = string("const_203_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 1, 512]> vp_cast_fp16 = pad(constant_val = const_203_to_fp16, mode = vp_mode_0, pad = vp_pad_0, x = var_7033_cast_fp16)[name = string("vp_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_7074_cast_fp16 = mul(x = K_c_cast_fp16, y = var_1005_cast_fp16)[name = string("op_7074_cast_fp16")];
+            tensor<int32, [4]> var_7075_reps_0 = const()[name = string("op_7075_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_7075_cast_fp16 = tile(reps = var_7075_reps_0, x = kp_cast_fp16)[name = string("op_7075_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_7076_cast_fp16 = mul(x = var_7075_cast_fp16, y = update_mask)[name = string("op_7076_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> K_n_cast_fp16 = add(x = var_7074_cast_fp16, y = var_7076_cast_fp16)[name = string("K_n_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_7082_cast_fp16 = mul(x = V_c_cast_fp16, y = var_1005_cast_fp16)[name = string("op_7082_cast_fp16")];
+            tensor<int32, [4]> var_7083_reps_0 = const()[name = string("op_7083_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
+            tensor<fp16, [1, 1, 512, 512]> var_7083_cast_fp16 = tile(reps = var_7083_reps_0, x = vp_cast_fp16)[name = string("op_7083_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> var_7084_cast_fp16 = mul(x = var_7083_cast_fp16, y = update_mask)[name = string("op_7084_cast_fp16")];
+            tensor<fp16, [1, 1, 512, 512]> V_n_cast_fp16 = add(x = var_7082_cast_fp16, y = var_7084_cast_fp16)[name = string("V_n_cast_fp16")];
+            tensor<int32, [1]> var_7088_axes_0 = const()[name = string("op_7088_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_7088_cast_fp16 = squeeze(axes = var_7088_axes_0, x = K_n_cast_fp16)[name = string("op_7088_cast_fp16")];
+            tensor<int32, [4]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_23_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_23_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_23_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_88, begin_mask = kv_cache_0_internal_tensor_assign_23_begin_mask_0, end = concat_89, end_mask = kv_cache_0_internal_tensor_assign_23_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_23_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_23_stride_0, update = var_7088_cast_fp16, x = coreml_update_state_45)[name = string("kv_cache_0_internal_tensor_assign_23_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_23_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_46_write_state")];
+            tensor<fp16, [24, 1, 512, 512]> coreml_update_state_46 = read_state(input = kv_cache_0)[name = string("coreml_update_state_46")];
+            tensor<int32, [1]> var_7095_axes_0 = const()[name = string("op_7095_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 512, 512]> var_7095_cast_fp16 = squeeze(axes = var_7095_axes_0, x = V_n_cast_fp16)[name = string("op_7095_cast_fp16")];
+            tensor<int32, [4]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> kv_cache_0_internal_tensor_assign_24_stride_0 = const()[name = string("kv_cache_0_internal_tensor_assign_24_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_24_begin_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_24_end_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> kv_cache_0_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("kv_cache_0_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [24, 1, 512, 512]> kv_cache_0_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_90, begin_mask = kv_cache_0_internal_tensor_assign_24_begin_mask_0, end = concat_91, end_mask = kv_cache_0_internal_tensor_assign_24_end_mask_0, squeeze_mask = kv_cache_0_internal_tensor_assign_24_squeeze_mask_0, stride = kv_cache_0_internal_tensor_assign_24_stride_0, update = var_7095_cast_fp16, x = coreml_update_state_46)[name = string("kv_cache_0_internal_tensor_assign_24_cast_fp16")];
+            write_state(data = kv_cache_0_internal_tensor_assign_24_cast_fp16, input = kv_cache_0)[name = string("coreml_update_state_47_write_state")];
+            tensor<int32, [4]> var_7105_begin_0 = const()[name = string("op_7105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_7105_end_0 = const()[name = string("op_7105_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_7105_end_mask_0 = const()[name = string("op_7105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_7105_cast_fp16 = slice_by_index(begin = var_7105_begin_0, end = var_7105_end_0, end_mask = var_7105_end_mask_0, x = K_n_cast_fp16)[name = string("op_7105_cast_fp16")];
+            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = var_7105_cast_fp16)[name = string("transpose_10")];
+            tensor<fp16, [8, 1, 512, 256]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
+            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_44_cast_fp16 = reshape(shape = concat_92, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
+            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_9")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_45_cast_fp16 = reshape(shape = concat_93, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
+            tensor<int32, [4]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
+            tensor<int32, [4]> var_7114_begin_0 = const()[name = string("op_7114_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_7114_end_0 = const()[name = string("op_7114_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
+            tensor<bool, [4]> var_7114_end_mask_0 = const()[name = string("op_7114_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 512, 256]> var_7114_cast_fp16 = slice_by_index(begin = var_7114_begin_0, end = var_7114_end_0, end_mask = var_7114_end_mask_0, x = V_n_cast_fp16)[name = string("op_7114_cast_fp16")];
+            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([8, 1, 1, 1])];
+            tensor<fp16, [1, 1, 512, 256]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = var_7114_cast_fp16)[name = string("transpose_8")];
+            tensor<fp16, [8, 1, 512, 256]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
+            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([8, 1, 1, 512, 256])];
+            tensor<fp16, [8, 1, 1, 512, 256]> reshape_46_cast_fp16 = reshape(shape = concat_94, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
+            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
+            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
+            tensor<fp16, [1, 8, 1, 512, 256]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_7")];
+            tensor<fp16, [8, 1, 512, 256]> reshape_47_cast_fp16 = reshape(shape = concat_95, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
+            tensor<int32, [4]> Ve_perm_0 = const()[name = string("Ve_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
+            bool var_7132_transpose_x_0 = const()[name = string("op_7132_transpose_x_0"), val = bool(false)];
+            bool var_7132_transpose_y_0 = const()[name = string("op_7132_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 256, 512]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_6")];
+            tensor<fp16, [1, 8, 1, 512]> var_7132_cast_fp16 = matmul(transpose_x = var_7132_transpose_x_0, transpose_y = var_7132_transpose_y_0, x = q, y = transpose_59_cast_fp16)[name = string("op_7132_cast_fp16")];
+            tensor<fp16, [1, 8, 1, 512]> var_7139_cast_fp16 = add(x = var_7132_cast_fp16, y = causal_mask)[name = string("op_7139_cast_fp16")];
+            int32 var_7140 = const()[name = string("op_7140"), val = int32(-1)];
+            tensor<fp16, [1, 8, 1, 512]> var_7142_cast_fp16 = softmax(axis = var_7140, x = var_7139_cast_fp16)[name = string("op_7142_cast_fp16")];
+            bool var_7158_transpose_x_0 = const()[name = string("op_7158_transpose_x_0"), val = bool(false)];
+            bool var_7158_transpose_y_0 = const()[name = string("op_7158_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 256]> Ve_cast_fp16 = transpose(perm = Ve_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_5")];
+            tensor<fp16, [1, 8, 1, 256]> var_7158_cast_fp16 = matmul(transpose_x = var_7158_transpose_x_0, transpose_y = var_7158_transpose_y_0, x = var_7142_cast_fp16, y = Ve_cast_fp16)[name = string("op_7158_cast_fp16")];
+            tensor<int32, [4]> var_7168 = const()[name = string("op_7168"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_7175 = const()[name = string("op_7175"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, 8, 256]> var_7169 = transpose(perm = var_7168, x = var_7158_cast_fp16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 2048]> var_7176 = reshape(shape = var_7175, x = var_7169)[name = string("op_7176")];
+            tensor<int32, [3]> var_7180 = const()[name = string("op_7180"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1536, 2048, 1]> squeeze_11_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223599232))), lut = tensor<fp16, [48, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225172160))))[name = string("squeeze_11_palettized")];
+            string var_7196_pad_type_0 = const()[name = string("op_7196_pad_type_0"), val = string("valid")];
+            int32 var_7196_groups_0 = const()[name = string("op_7196_groups_0"), val = int32(1)];
+            tensor<int32, [1]> var_7196_strides_0 = const()[name = string("op_7196_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_7196_pad_0 = const()[name = string("op_7196_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_7196_dilations_0 = const()[name = string("op_7196_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 2048, 1]> var_7181 = transpose(perm = var_7180, x = var_7176)[name = string("transpose_3")];
+            tensor<fp16, [1, 1536, 1]> var_7196 = conv(dilations = var_7196_dilations_0, groups = var_7196_groups_0, pad = var_7196_pad_0, pad_type = var_7196_pad_type_0, strides = var_7196_strides_0, weight = squeeze_11_palettized, x = var_7181)[name = string("op_7196")];
+            tensor<int32, [3]> var_7200 = const()[name = string("op_7200"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_7206 = const()[name = string("op_7206"), val = int32(-1)];
+            fp16 const_204_promoted_to_fp16 = const()[name = string("const_204_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_343 = transpose(perm = var_7200, x = var_7196)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 1536]> var_7212_cast_fp16 = mul(x = x_343, y = const_204_promoted_to_fp16)[name = string("op_7212_cast_fp16")];
+            bool input_337_interleave_0 = const()[name = string("input_337_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_337_cast_fp16 = concat(axis = var_7206, interleave = input_337_interleave_0, values = (x_343, var_7212_cast_fp16))[name = string("input_337_cast_fp16")];
+            tensor<int32, [1]> normed_321_axes_0 = const()[name = string("normed_321_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_321_cast_fp16 = layer_norm(axes = normed_321_axes_0, epsilon = var_7204_to_fp16, x = input_337_cast_fp16)[name = string("normed_321_cast_fp16")];
+            tensor<int32, [2]> var_7217_split_sizes_0 = const()[name = string("op_7217_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_7217_axis_0 = const()[name = string("op_7217_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_7217_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_7217_cast_fp16_1 = split(axis = var_7217_axis_0, split_sizes = var_7217_split_sizes_0, x = normed_321_cast_fp16)[name = string("op_7217_cast_fp16")];
+            tensor<fp16, [1536]> const_205_to_fp16 = const()[name = string("const_205_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225173760)))];
+            tensor<fp16, [1, 1, 1536]> var_7220_cast_fp16 = mul(x = var_7217_cast_fp16_0, y = const_205_to_fp16)[name = string("op_7220_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> x_347_cast_fp16 = add(x = x_329_cast_fp16, y = var_7220_cast_fp16)[name = string("x_347_cast_fp16")];
+            int32 var_7227 = const()[name = string("op_7227"), val = int32(-1)];
+            fp16 const_206_promoted_to_fp16 = const()[name = string("const_206_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_7233_cast_fp16 = mul(x = x_347_cast_fp16, y = const_206_promoted_to_fp16)[name = string("op_7233_cast_fp16")];
+            bool input_339_interleave_0 = const()[name = string("input_339_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_339_cast_fp16 = concat(axis = var_7227, interleave = input_339_interleave_0, values = (x_347_cast_fp16, var_7233_cast_fp16))[name = string("input_339_cast_fp16")];
+            tensor<int32, [1]> normed_325_axes_0 = const()[name = string("normed_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_7225_to_fp16 = const()[name = string("op_7225_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_325_cast_fp16 = layer_norm(axes = normed_325_axes_0, epsilon = var_7225_to_fp16, x = input_339_cast_fp16)[name = string("normed_325_cast_fp16")];
+            tensor<int32, [2]> var_7238_split_sizes_0 = const()[name = string("op_7238_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_7238_axis_0 = const()[name = string("op_7238_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_7238_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_7238_cast_fp16_1 = split(axis = var_7238_axis_0, split_sizes = var_7238_split_sizes_0, x = normed_325_cast_fp16)[name = string("op_7238_cast_fp16")];
+            tensor<fp16, [1536]> const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225176896)))];
+            tensor<fp16, [1, 1, 1536]> var_7241_cast_fp16 = mul(x = var_7238_cast_fp16_0, y = const_207_to_fp16)[name = string("op_7241_cast_fp16")];
+            tensor<int32, [3]> var_7254 = const()[name = string("op_7254"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> input_341_axes_0 = const()[name = string("input_341_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_7255 = transpose(perm = var_7254, x = var_7241_cast_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1536, 1, 1]> input_341 = expand_dims(axes = input_341_axes_0, x = var_7255)[name = string("input_341")];
+            string var_7268_pad_type_0 = const()[name = string("op_7268_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_7268_strides_0 = const()[name = string("op_7268_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_7268_pad_0 = const()[name = string("op_7268_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_7268_dilations_0 = const()[name = string("op_7268_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_7268_groups_0 = const()[name = string("op_7268_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_7268 = conv(dilations = var_7268_dilations_0, groups = var_7268_groups_0, pad = var_7268_pad_0, pad_type = var_7268_pad_type_0, strides = var_7268_strides_0, weight = layers_11_mlp_gate_proj_weight_palettized, x = input_341)[name = string("op_7268")];
+            string var_7270_mode_0 = const()[name = string("op_7270_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 6144, 1, 1]> var_7270 = gelu(mode = var_7270_mode_0, x = var_7268)[name = string("op_7270")];
+            string var_7281_pad_type_0 = const()[name = string("op_7281_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_7281_strides_0 = const()[name = string("op_7281_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_7281_pad_0 = const()[name = string("op_7281_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_7281_dilations_0 = const()[name = string("op_7281_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_7281_groups_0 = const()[name = string("op_7281_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 6144, 1, 1]> var_7281 = conv(dilations = var_7281_dilations_0, groups = var_7281_groups_0, pad = var_7281_pad_0, pad_type = var_7281_pad_type_0, strides = var_7281_strides_0, weight = layers_11_mlp_up_proj_weight_palettized, x = input_341)[name = string("op_7281")];
+            tensor<fp16, [1, 6144, 1, 1]> input_343 = mul(x = var_7270, y = var_7281)[name = string("input_343")];
+            string var_7293_pad_type_0 = const()[name = string("op_7293_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> var_7293_strides_0 = const()[name = string("op_7293_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_7293_pad_0 = const()[name = string("op_7293_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_7293_dilations_0 = const()[name = string("op_7293_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_7293_groups_0 = const()[name = string("op_7293_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 1536, 1, 1]> var_7293 = conv(dilations = var_7293_dilations_0, groups = var_7293_groups_0, pad = var_7293_pad_0, pad_type = var_7293_pad_type_0, strides = var_7293_strides_0, weight = layers_11_mlp_down_proj_weight_palettized, x = input_343)[name = string("op_7293")];
+            tensor<int32, [1]> var_7295_axes_0 = const()[name = string("op_7295_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1536, 1]> var_7295 = squeeze(axes = var_7295_axes_0, x = var_7293)[name = string("op_7295")];
+            tensor<int32, [3]> var_7299 = const()[name = string("op_7299"), val = tensor<int32, [3]>([0, 2, 1])];
+            int32 var_7305 = const()[name = string("op_7305"), val = int32(-1)];
+            fp16 const_208_promoted_to_fp16 = const()[name = string("const_208_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> x_351 = transpose(perm = var_7299, x = var_7295)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 1536]> var_7311_cast_fp16 = mul(x = x_351, y = const_208_promoted_to_fp16)[name = string("op_7311_cast_fp16")];
+            bool input_345_interleave_0 = const()[name = string("input_345_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_345_cast_fp16 = concat(axis = var_7305, interleave = input_345_interleave_0, values = (x_351, var_7311_cast_fp16))[name = string("input_345_cast_fp16")];
+            tensor<int32, [1]> normed_329_axes_0 = const()[name = string("normed_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_7303_to_fp16 = const()[name = string("op_7303_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_329_cast_fp16 = layer_norm(axes = normed_329_axes_0, epsilon = var_7303_to_fp16, x = input_345_cast_fp16)[name = string("normed_329_cast_fp16")];
+            tensor<int32, [2]> var_7316_split_sizes_0 = const()[name = string("op_7316_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_7316_axis_0 = const()[name = string("op_7316_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_7316_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_7316_cast_fp16_1 = split(axis = var_7316_axis_0, split_sizes = var_7316_split_sizes_0, x = normed_329_cast_fp16)[name = string("op_7316_cast_fp16")];
+            tensor<fp16, [1536]> const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225180032)))];
+            tensor<fp16, [1, 1, 1536]> var_7319_cast_fp16 = mul(x = var_7316_cast_fp16_0, y = const_209_to_fp16)[name = string("op_7319_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_167_cast_fp16 = add(x = x_347_cast_fp16, y = var_7319_cast_fp16)[name = string("hidden_states_167_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> var_7330 = linear(bias = linear_0_bias_0, weight = layers_11_per_layer_input_gate_weight_palettized, x = hidden_states_167_cast_fp16)[name = string("linear_22")];
+            string gated_mode_0 = const()[name = string("gated_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 1, 256]> gated = gelu(mode = gated_mode_0, x = var_7330)[name = string("gated")];
+            tensor<int32, [3]> var_7347_begin_0 = const()[name = string("op_7347_begin_0"), val = tensor<int32, [3]>([0, 0, 2816])];
+            tensor<int32, [3]> var_7347_end_0 = const()[name = string("op_7347_end_0"), val = tensor<int32, [3]>([1, 1, 3072])];
+            tensor<bool, [3]> var_7347_end_mask_0 = const()[name = string("op_7347_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 1, 256]> var_7347_cast_fp16 = slice_by_index(begin = var_7347_begin_0, end = var_7347_end_0, end_mask = var_7347_end_mask_0, x = per_layer_combined)[name = string("op_7347_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> input_349_cast_fp16 = mul(x = gated, y = var_7347_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<fp16, [1536, 256]> layers_11_per_layer_projection_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1536, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225183168))), lut = tensor<fp16, [48, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225379840))))[name = string("layers_11_per_layer_projection_weight_promoted_to_fp16_palettized")];
+            tensor<fp16, [1, 1, 1536]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_per_layer_projection_weight_promoted_to_fp16_palettized, x = input_349_cast_fp16)[name = string("linear_23_cast_fp16")];
+            int32 var_7356 = const()[name = string("op_7356"), val = int32(-1)];
+            fp16 const_210_promoted_to_fp16 = const()[name = string("const_210_promoted_to_fp16"), val = fp16(-0x1p+0)];
+            tensor<fp16, [1, 1, 1536]> var_7362_cast_fp16 = mul(x = linear_23_cast_fp16, y = const_210_promoted_to_fp16)[name = string("op_7362_cast_fp16")];
+            bool input_interleave_0 = const()[name = string("input_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 3072]> input_cast_fp16 = concat(axis = var_7356, interleave = input_interleave_0, values = (linear_23_cast_fp16, var_7362_cast_fp16))[name = string("input_cast_fp16")];
+            tensor<int32, [1]> normed_333_axes_0 = const()[name = string("normed_333_axes_0"), val = tensor<int32, [1]>([-1])];
+            fp16 var_7354_to_fp16 = const()[name = string("op_7354_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 1, 3072]> normed_333_cast_fp16 = layer_norm(axes = normed_333_axes_0, epsilon = var_7354_to_fp16, x = input_cast_fp16)[name = string("normed_333_cast_fp16")];
+            tensor<int32, [2]> var_7367_split_sizes_0 = const()[name = string("op_7367_split_sizes_0"), val = tensor<int32, [2]>([1536, 1536])];
+            int32 var_7367_axis_0 = const()[name = string("op_7367_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1536]> var_7367_cast_fp16_0, tensor<fp16, [1, 1, 1536]> var_7367_cast_fp16_1 = split(axis = var_7367_axis_0, split_sizes = var_7367_split_sizes_0, x = normed_333_cast_fp16)[name = string("op_7367_cast_fp16")];
+            tensor<fp16, [1536]> const_211_to_fp16 = const()[name = string("const_211_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225381440)))];
+            tensor<fp16, [1, 1, 1536]> var_7370_cast_fp16 = mul(x = var_7367_cast_fp16_0, y = const_211_to_fp16)[name = string("op_7370_cast_fp16")];
+            tensor<fp16, [1, 1, 1536]> hidden_states_cast_fp16 = add(x = hidden_states_167_cast_fp16, y = var_7370_cast_fp16)[name = string("hidden_states_cast_fp16")];
+            tensor<fp16, [1]> layers_11_layer_scalar_to_fp16 = const()[name = string("layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-2])];
+            tensor<fp16, [1, 1, 1536]> hidden_states_out = mul(x = hidden_states_cast_fp16, y = layers_11_layer_scalar_to_fp16)[name = string("op_7374_cast_fp16")];
+        } -> (hidden_states_out);
+}
\ No newline at end of file