openai_whisper-large-v3-v20240930_547MB/TextDecoder.mlmodelc/model.mil
263.3 KB · 1300 lines · plaintext Raw
1 program(1.0)
2 [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}})]
3 {
4 func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 5120, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 5120, 1, 448]> value_cache) {
5 tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)];
6 tensor<int32, []> var_24_batch_dims_0 = const()[name = tensor<string, []>("op_24_batch_dims_0"), val = tensor<int32, []>(0)];
7 tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
8 tensor<fp16, [1, 1280]> var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
9 tensor<int32, []> var_28_axis_0 = const()[name = tensor<string, []>("op_28_axis_0"), val = tensor<int32, []>(0)];
10 tensor<int32, []> var_28_batch_dims_0 = const()[name = tensor<string, []>("op_28_batch_dims_0"), val = tensor<int32, []>(0)];
11 tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132777088)))];
12 tensor<fp16, [1, 1280]> var_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
13 tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
14 tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])];
15 tensor<fp16, [1, 1280, 1]> var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_42_cast_fp16")];
16 tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
17 tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_42_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
18 tensor<int32, [4]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])];
19 tensor<int32, []> var_47_axis_0 = const()[name = tensor<string, []>("op_47_axis_0"), val = tensor<int32, []>(1)];
20 tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_47_cast_fp16")];
21 tensor<int32, [4]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])];
22 tensor<int32, []> var_54_axis_0 = const()[name = tensor<string, []>("op_54_axis_0"), val = tensor<int32, []>(1)];
23 tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_54_cast_fp16")];
24 tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(3)];
25 tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
26 tensor<fp16, []> var_90_to_fp16 = const()[name = tensor<string, []>("op_90_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
27 tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_90_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
28 tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133924032)))];
29 tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133926656)))];
30 tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133929280)))];
31 tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133931904)))];
32 tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
33 tensor<fp16, [1, 1280, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
34 tensor<string, []> pretrained_out_1_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_1_pad_type_0"), val = tensor<string, []>("valid")];
35 tensor<int32, [2]> pretrained_out_1_strides_0 = const()[name = tensor<string, []>("pretrained_out_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
36 tensor<int32, [4]> pretrained_out_1_pad_0 = const()[name = tensor<string, []>("pretrained_out_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
37 tensor<int32, [2]> pretrained_out_1_dilations_0 = const()[name = tensor<string, []>("pretrained_out_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
38 tensor<int32, []> pretrained_out_1_groups_0 = const()[name = tensor<string, []>("pretrained_out_1_groups_0"), val = tensor<int32, []>(1)];
39 tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133934528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134753792))), name = tensor<string, []>("layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
40 tensor<fp16, [1280]> layers_0_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134753920)))];
41 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_1_dilations_0, groups = pretrained_out_1_groups_0, pad = pretrained_out_1_pad_0, pad_type = pretrained_out_1_pad_type_0, strides = pretrained_out_1_strides_0, weight = layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("pretrained_out_1_cast_fp16")];
42 tensor<string, []> input_1_pad_type_0 = const()[name = tensor<string, []>("input_1_pad_type_0"), val = tensor<string, []>("valid")];
43 tensor<int32, [2]> input_1_strides_0 = const()[name = tensor<string, []>("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
44 tensor<int32, [4]> input_1_pad_0 = const()[name = tensor<string, []>("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
45 tensor<int32, [2]> input_1_dilations_0 = const()[name = tensor<string, []>("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
46 tensor<int32, []> input_1_groups_0 = const()[name = tensor<string, []>("input_1_groups_0"), val = tensor<int32, []>(1)];
47 tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134756544)))];
48 tensor<fp16, [1, 16, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_self_attn_q_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
49 tensor<string, []> lora_out_1_pad_type_0 = const()[name = tensor<string, []>("lora_out_1_pad_type_0"), val = tensor<string, []>("valid")];
50 tensor<int32, [2]> lora_out_1_strides_0 = const()[name = tensor<string, []>("lora_out_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
51 tensor<int32, [4]> lora_out_1_pad_0 = const()[name = tensor<string, []>("lora_out_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
52 tensor<int32, [2]> lora_out_1_dilations_0 = const()[name = tensor<string, []>("lora_out_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
53 tensor<int32, []> lora_out_1_groups_0 = const()[name = tensor<string, []>("lora_out_1_groups_0"), val = tensor<int32, []>(1)];
54 tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134797568)))];
55 tensor<fp16, [1, 1280, 1, 1]> lora_out_1_cast_fp16 = conv(dilations = lora_out_1_dilations_0, groups = lora_out_1_groups_0, pad = lora_out_1_pad_0, pad_type = lora_out_1_pad_type_0, strides = lora_out_1_strides_0, weight = layers_0_self_attn_q_proj_loraB_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("lora_out_1_cast_fp16")];
56 tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = add(x = pretrained_out_1_cast_fp16, y = lora_out_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
57 tensor<string, []> pretrained_out_3_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_3_pad_type_0"), val = tensor<string, []>("valid")];
58 tensor<int32, [2]> pretrained_out_3_strides_0 = const()[name = tensor<string, []>("pretrained_out_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
59 tensor<int32, [4]> pretrained_out_3_pad_0 = const()[name = tensor<string, []>("pretrained_out_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
60 tensor<int32, [2]> pretrained_out_3_dilations_0 = const()[name = tensor<string, []>("pretrained_out_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
61 tensor<int32, []> pretrained_out_3_groups_0 = const()[name = tensor<string, []>("pretrained_out_3_groups_0"), val = tensor<int32, []>(1)];
62 tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134838592))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135657856))), name = tensor<string, []>("layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
63 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_3_cast_fp16 = conv(dilations = pretrained_out_3_dilations_0, groups = pretrained_out_3_groups_0, pad = pretrained_out_3_pad_0, pad_type = pretrained_out_3_pad_type_0, strides = pretrained_out_3_strides_0, weight = layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("pretrained_out_3_cast_fp16")];
64 tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("valid")];
65 tensor<int32, [2]> input_3_strides_0 = const()[name = tensor<string, []>("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
66 tensor<int32, [4]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
67 tensor<int32, [2]> input_3_dilations_0 = const()[name = tensor<string, []>("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
68 tensor<int32, []> input_3_groups_0 = const()[name = tensor<string, []>("input_3_groups_0"), val = tensor<int32, []>(1)];
69 tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135657984)))];
70 tensor<fp16, [1, 16, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_0_self_attn_k_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
71 tensor<string, []> lora_out_3_pad_type_0 = const()[name = tensor<string, []>("lora_out_3_pad_type_0"), val = tensor<string, []>("valid")];
72 tensor<int32, [2]> lora_out_3_strides_0 = const()[name = tensor<string, []>("lora_out_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
73 tensor<int32, [4]> lora_out_3_pad_0 = const()[name = tensor<string, []>("lora_out_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
74 tensor<int32, [2]> lora_out_3_dilations_0 = const()[name = tensor<string, []>("lora_out_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
75 tensor<int32, []> lora_out_3_groups_0 = const()[name = tensor<string, []>("lora_out_3_groups_0"), val = tensor<int32, []>(1)];
76 tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135699008)))];
77 tensor<fp16, [1, 1280, 1, 1]> lora_out_3_cast_fp16 = conv(dilations = lora_out_3_dilations_0, groups = lora_out_3_groups_0, pad = lora_out_3_pad_0, pad_type = lora_out_3_pad_type_0, strides = lora_out_3_strides_0, weight = layers_0_self_attn_k_proj_loraB_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("lora_out_3_cast_fp16")];
78 tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = add(x = pretrained_out_3_cast_fp16, y = lora_out_3_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")];
79 tensor<string, []> pretrained_out_5_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_5_pad_type_0"), val = tensor<string, []>("valid")];
80 tensor<int32, [2]> pretrained_out_5_strides_0 = const()[name = tensor<string, []>("pretrained_out_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
81 tensor<int32, [4]> pretrained_out_5_pad_0 = const()[name = tensor<string, []>("pretrained_out_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
82 tensor<int32, [2]> pretrained_out_5_dilations_0 = const()[name = tensor<string, []>("pretrained_out_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
83 tensor<int32, []> pretrained_out_5_groups_0 = const()[name = tensor<string, []>("pretrained_out_5_groups_0"), val = tensor<int32, []>(1)];
84 tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135740032))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136559296))), name = tensor<string, []>("layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
85 tensor<fp16, [1280]> layers_0_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136559424)))];
86 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_5_dilations_0, groups = pretrained_out_5_groups_0, pad = pretrained_out_5_pad_0, pad_type = pretrained_out_5_pad_type_0, strides = pretrained_out_5_strides_0, weight = layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("pretrained_out_5_cast_fp16")];
87 tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")];
88 tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
89 tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
90 tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
91 tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)];
92 tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136562048)))];
93 tensor<fp16, [1, 16, 1, 1]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_0_self_attn_v_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
94 tensor<string, []> lora_out_5_pad_type_0 = const()[name = tensor<string, []>("lora_out_5_pad_type_0"), val = tensor<string, []>("valid")];
95 tensor<int32, [2]> lora_out_5_strides_0 = const()[name = tensor<string, []>("lora_out_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
96 tensor<int32, [4]> lora_out_5_pad_0 = const()[name = tensor<string, []>("lora_out_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
97 tensor<int32, [2]> lora_out_5_dilations_0 = const()[name = tensor<string, []>("lora_out_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
98 tensor<int32, []> lora_out_5_groups_0 = const()[name = tensor<string, []>("lora_out_5_groups_0"), val = tensor<int32, []>(1)];
99 tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136603072)))];
100 tensor<fp16, [1, 1280, 1, 1]> lora_out_5_cast_fp16 = conv(dilations = lora_out_5_dilations_0, groups = lora_out_5_groups_0, pad = lora_out_5_pad_0, pad_type = lora_out_5_pad_type_0, strides = lora_out_5_strides_0, weight = layers_0_self_attn_v_proj_loraB_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("lora_out_5_cast_fp16")];
101 tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = add(x = pretrained_out_5_cast_fp16, y = lora_out_5_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")];
102 tensor<int32, [1]> var_173_axes_0 = const()[name = tensor<string, []>("op_173_axes_0"), val = tensor<int32, [1]>([1])];
103 tensor<fp16, [1, 1, 448]> var_173_cast_fp16 = expand_dims(axes = var_173_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_173_cast_fp16")];
104 tensor<int32, [1]> var_174_axes_0 = const()[name = tensor<string, []>("op_174_axes_0"), val = tensor<int32, [1]>([2])];
105 tensor<fp16, [1, 1, 1, 448]> var_174_cast_fp16 = expand_dims(axes = var_174_axes_0, x = var_173_cast_fp16)[name = tensor<string, []>("op_174_cast_fp16")];
106 tensor<fp16, [1, 1280, 1, 448]> var_176_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_176_cast_fp16")];
107 tensor<fp16, []> var_65_to_fp16 = const()[name = tensor<string, []>("op_65_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
108 tensor<fp16, [1, 1, 1, 448]> var_177_cast_fp16 = sub(x = var_65_to_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_177_cast_fp16")];
109 tensor<fp16, [1, 1280, 1, 448]> var_178_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_177_cast_fp16)[name = tensor<string, []>("op_178_cast_fp16")];
110 tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_176_cast_fp16, y = var_178_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
111 tensor<fp16, [1, 1280, 1, 448]> var_180_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_180_cast_fp16")];
112 tensor<fp16, [1, 1280, 1, 448]> var_182_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_177_cast_fp16)[name = tensor<string, []>("op_182_cast_fp16")];
113 tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_180_cast_fp16, y = var_182_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
114 tensor<int32, [4]> var_185 = const()[name = tensor<string, []>("op_185"), val = tensor<int32, [4]>([1, 20, 64, -1])];
115 tensor<fp16, [1, 20, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_185, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
116 tensor<fp16, []> var_187_to_fp16 = const()[name = tensor<string, []>("op_187_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
117 tensor<fp16, [1, 20, 64, 1]> var_188_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_187_to_fp16)[name = tensor<string, []>("op_188_cast_fp16")];
118 tensor<int32, [4]> var_189 = const()[name = tensor<string, []>("op_189"), val = tensor<int32, [4]>([1, 20, 64, -1])];
119 tensor<fp16, [1, 20, 64, 448]> var_190_cast_fp16 = reshape(shape = var_189, x = key_1_cast_fp16)[name = tensor<string, []>("op_190_cast_fp16")];
120 tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
121 tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
122 tensor<fp16, [1, 20, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_188_cast_fp16, y = var_190_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
123 tensor<int32, [1]> var_194_axes_0 = const()[name = tensor<string, []>("op_194_axes_0"), val = tensor<int32, [1]>([1])];
124 tensor<fp16, [1, 1, 448]> var_194_cast_fp16 = expand_dims(axes = var_194_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_194_cast_fp16")];
125 tensor<int32, [1]> var_195_axes_0 = const()[name = tensor<string, []>("op_195_axes_0"), val = tensor<int32, [1]>([2])];
126 tensor<fp16, [1, 1, 1, 448]> var_195_cast_fp16 = expand_dims(axes = var_195_axes_0, x = var_194_cast_fp16)[name = tensor<string, []>("op_195_cast_fp16")];
127 tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
128 tensor<fp16, [1, 20, 1, 448]> var_198_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
129 tensor<int32, [4]> var_199 = const()[name = tensor<string, []>("op_199"), val = tensor<int32, [4]>([1, 20, 64, -1])];
130 tensor<fp16, [1, 20, 64, 448]> var_200_cast_fp16 = reshape(shape = var_199, x = value_1_cast_fp16)[name = tensor<string, []>("op_200_cast_fp16")];
131 tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
132 tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
133 tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_200_cast_fp16, y = var_198_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
134 tensor<int32, [4]> var_203 = const()[name = tensor<string, []>("op_203"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
135 tensor<fp16, [1, 1280, 1, 1]> input_7_cast_fp16 = reshape(shape = var_203, x = attn_1_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
136 tensor<string, []> pretrained_out_7_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_7_pad_type_0"), val = tensor<string, []>("valid")];
137 tensor<int32, [2]> pretrained_out_7_strides_0 = const()[name = tensor<string, []>("pretrained_out_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
138 tensor<int32, [4]> pretrained_out_7_pad_0 = const()[name = tensor<string, []>("pretrained_out_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
139 tensor<int32, [2]> pretrained_out_7_dilations_0 = const()[name = tensor<string, []>("pretrained_out_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
140 tensor<int32, []> pretrained_out_7_groups_0 = const()[name = tensor<string, []>("pretrained_out_7_groups_0"), val = tensor<int32, []>(1)];
141 tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136644096))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137463360))), name = tensor<string, []>("layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
142 tensor<fp16, [1280]> layers_0_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137463488)))];
143 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_7_dilations_0, groups = pretrained_out_7_groups_0, pad = pretrained_out_7_pad_0, pad_type = pretrained_out_7_pad_type_0, strides = pretrained_out_7_strides_0, weight = layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("pretrained_out_7_cast_fp16")];
144 tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("valid")];
145 tensor<int32, [2]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
146 tensor<int32, [4]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
147 tensor<int32, [2]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
148 tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)];
149 tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137466112)))];
150 tensor<fp16, [1, 16, 1, 1]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_0_self_attn_o_proj_loraA_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
151 tensor<string, []> lora_out_7_pad_type_0 = const()[name = tensor<string, []>("lora_out_7_pad_type_0"), val = tensor<string, []>("valid")];
152 tensor<int32, [2]> lora_out_7_strides_0 = const()[name = tensor<string, []>("lora_out_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
153 tensor<int32, [4]> lora_out_7_pad_0 = const()[name = tensor<string, []>("lora_out_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
154 tensor<int32, [2]> lora_out_7_dilations_0 = const()[name = tensor<string, []>("lora_out_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
155 tensor<int32, []> lora_out_7_groups_0 = const()[name = tensor<string, []>("lora_out_7_groups_0"), val = tensor<int32, []>(1)];
156 tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137507136)))];
157 tensor<fp16, [1, 1280, 1, 1]> lora_out_7_cast_fp16 = conv(dilations = lora_out_7_dilations_0, groups = lora_out_7_groups_0, pad = lora_out_7_pad_0, pad_type = lora_out_7_pad_type_0, strides = lora_out_7_strides_0, weight = layers_0_self_attn_o_proj_loraB_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("lora_out_7_cast_fp16")];
158 tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = add(x = pretrained_out_7_cast_fp16, y = lora_out_7_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
159 tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
160 tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
161 tensor<fp16, []> var_241_to_fp16 = const()[name = tensor<string, []>("op_241_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
162 tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_241_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
163 tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137548160)))];
164 tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137550784)))];
165 tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
166 tensor<fp16, [1, 1280, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
167 tensor<string, []> pretrained_out_9_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_9_pad_type_0"), val = tensor<string, []>("valid")];
168 tensor<int32, [2]> pretrained_out_9_strides_0 = const()[name = tensor<string, []>("pretrained_out_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
169 tensor<int32, [4]> pretrained_out_9_pad_0 = const()[name = tensor<string, []>("pretrained_out_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
170 tensor<int32, [2]> pretrained_out_9_dilations_0 = const()[name = tensor<string, []>("pretrained_out_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
171 tensor<int32, []> pretrained_out_9_groups_0 = const()[name = tensor<string, []>("pretrained_out_9_groups_0"), val = tensor<int32, []>(1)];
172 tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137553408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138372672))), name = tensor<string, []>("layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
173 tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138372800)))];
174 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_9_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_9_dilations_0, groups = pretrained_out_9_groups_0, pad = pretrained_out_9_pad_0, pad_type = pretrained_out_9_pad_type_0, strides = pretrained_out_9_strides_0, weight = layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("pretrained_out_9_cast_fp16")];
175 tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")];
176 tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
177 tensor<int32, [4]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
178 tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
179 tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(1)];
180 tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138375424)))];
181 tensor<fp16, [1, 16, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_0_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
182 tensor<string, []> lora_out_9_pad_type_0 = const()[name = tensor<string, []>("lora_out_9_pad_type_0"), val = tensor<string, []>("valid")];
183 tensor<int32, [2]> lora_out_9_strides_0 = const()[name = tensor<string, []>("lora_out_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
184 tensor<int32, [4]> lora_out_9_pad_0 = const()[name = tensor<string, []>("lora_out_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
185 tensor<int32, [2]> lora_out_9_dilations_0 = const()[name = tensor<string, []>("lora_out_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
186 tensor<int32, []> lora_out_9_groups_0 = const()[name = tensor<string, []>("lora_out_9_groups_0"), val = tensor<int32, []>(1)];
187 tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138416448)))];
188 tensor<fp16, [1, 1280, 1, 1]> lora_out_9_cast_fp16 = conv(dilations = lora_out_9_dilations_0, groups = lora_out_9_groups_0, pad = lora_out_9_pad_0, pad_type = lora_out_9_pad_type_0, strides = lora_out_9_strides_0, weight = layers_0_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("lora_out_9_cast_fp16")];
189 tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = add(x = pretrained_out_9_cast_fp16, y = lora_out_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
190 tensor<string, []> pretrained_out_11_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_11_pad_type_0"), val = tensor<string, []>("valid")];
191 tensor<int32, [2]> pretrained_out_11_strides_0 = const()[name = tensor<string, []>("pretrained_out_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
192 tensor<int32, [4]> pretrained_out_11_pad_0 = const()[name = tensor<string, []>("pretrained_out_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
193 tensor<int32, [2]> pretrained_out_11_dilations_0 = const()[name = tensor<string, []>("pretrained_out_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
194 tensor<int32, []> pretrained_out_11_groups_0 = const()[name = tensor<string, []>("pretrained_out_11_groups_0"), val = tensor<int32, []>(1)];
195 tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138457472))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139276736))), name = tensor<string, []>("layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
196 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_11_cast_fp16 = conv(dilations = pretrained_out_11_dilations_0, groups = pretrained_out_11_groups_0, pad = pretrained_out_11_pad_0, pad_type = pretrained_out_11_pad_type_0, strides = pretrained_out_11_strides_0, weight = layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_11_cast_fp16")];
197 tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
198 tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
199 tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
200 tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
201 tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
202 tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139276864)))];
203 tensor<fp16, [1, 16, 1, 1500]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_0_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_13_cast_fp16")];
204 tensor<string, []> lora_out_11_pad_type_0 = const()[name = tensor<string, []>("lora_out_11_pad_type_0"), val = tensor<string, []>("valid")];
205 tensor<int32, [2]> lora_out_11_strides_0 = const()[name = tensor<string, []>("lora_out_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
206 tensor<int32, [4]> lora_out_11_pad_0 = const()[name = tensor<string, []>("lora_out_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
207 tensor<int32, [2]> lora_out_11_dilations_0 = const()[name = tensor<string, []>("lora_out_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
208 tensor<int32, []> lora_out_11_groups_0 = const()[name = tensor<string, []>("lora_out_11_groups_0"), val = tensor<int32, []>(1)];
209 tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139317888)))];
210 tensor<fp16, [1, 1280, 1, 1500]> lora_out_11_cast_fp16 = conv(dilations = lora_out_11_dilations_0, groups = lora_out_11_groups_0, pad = lora_out_11_pad_0, pad_type = lora_out_11_pad_type_0, strides = lora_out_11_strides_0, weight = layers_0_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("lora_out_11_cast_fp16")];
211 tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = add(x = pretrained_out_11_cast_fp16, y = lora_out_11_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
212 tensor<string, []> pretrained_out_13_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_13_pad_type_0"), val = tensor<string, []>("valid")];
213 tensor<int32, [2]> pretrained_out_13_strides_0 = const()[name = tensor<string, []>("pretrained_out_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
214 tensor<int32, [4]> pretrained_out_13_pad_0 = const()[name = tensor<string, []>("pretrained_out_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
215 tensor<int32, [2]> pretrained_out_13_dilations_0 = const()[name = tensor<string, []>("pretrained_out_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
216 tensor<int32, []> pretrained_out_13_groups_0 = const()[name = tensor<string, []>("pretrained_out_13_groups_0"), val = tensor<int32, []>(1)];
217 tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139358912))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140178176))), name = tensor<string, []>("layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
218 tensor<fp16, [1280]> layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140178304)))];
219 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_13_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_13_dilations_0, groups = pretrained_out_13_groups_0, pad = pretrained_out_13_pad_0, pad_type = pretrained_out_13_pad_type_0, strides = pretrained_out_13_strides_0, weight = layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_13_cast_fp16")];
220 tensor<string, []> input_15_pad_type_0 = const()[name = tensor<string, []>("input_15_pad_type_0"), val = tensor<string, []>("valid")];
221 tensor<int32, [2]> input_15_strides_0 = const()[name = tensor<string, []>("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
222 tensor<int32, [4]> input_15_pad_0 = const()[name = tensor<string, []>("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
223 tensor<int32, [2]> input_15_dilations_0 = const()[name = tensor<string, []>("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
224 tensor<int32, []> input_15_groups_0 = const()[name = tensor<string, []>("input_15_groups_0"), val = tensor<int32, []>(1)];
225 tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140180928)))];
226 tensor<fp16, [1, 16, 1, 1500]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_0_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_15_cast_fp16")];
227 tensor<string, []> lora_out_13_pad_type_0 = const()[name = tensor<string, []>("lora_out_13_pad_type_0"), val = tensor<string, []>("valid")];
228 tensor<int32, [2]> lora_out_13_strides_0 = const()[name = tensor<string, []>("lora_out_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
229 tensor<int32, [4]> lora_out_13_pad_0 = const()[name = tensor<string, []>("lora_out_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
230 tensor<int32, [2]> lora_out_13_dilations_0 = const()[name = tensor<string, []>("lora_out_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
231 tensor<int32, []> lora_out_13_groups_0 = const()[name = tensor<string, []>("lora_out_13_groups_0"), val = tensor<int32, []>(1)];
232 tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140221952)))];
233 tensor<fp16, [1, 1280, 1, 1500]> lora_out_13_cast_fp16 = conv(dilations = lora_out_13_dilations_0, groups = lora_out_13_groups_0, pad = lora_out_13_pad_0, pad_type = lora_out_13_pad_type_0, strides = lora_out_13_strides_0, weight = layers_0_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("lora_out_13_cast_fp16")];
234 tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = add(x = pretrained_out_13_cast_fp16, y = lora_out_13_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
235 tensor<int32, [4]> var_324 = const()[name = tensor<string, []>("op_324"), val = tensor<int32, [4]>([1, 20, 64, -1])];
236 tensor<fp16, [1, 20, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_324, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
237 tensor<fp16, []> var_326_to_fp16 = const()[name = tensor<string, []>("op_326_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
238 tensor<fp16, [1, 20, 64, 1]> var_327_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_326_to_fp16)[name = tensor<string, []>("op_327_cast_fp16")];
239 tensor<int32, [4]> var_328 = const()[name = tensor<string, []>("op_328"), val = tensor<int32, [4]>([1, 20, 64, -1])];
240 tensor<fp16, [1, 20, 64, 1500]> var_329_cast_fp16 = reshape(shape = var_328, x = key_3_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")];
241 tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
242 tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
243 tensor<fp16, [1, 20, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_327_cast_fp16, y = var_329_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
244 tensor<fp16, [1, 20, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
245 tensor<int32, [4]> var_333 = const()[name = tensor<string, []>("op_333"), val = tensor<int32, [4]>([1, 20, 64, -1])];
246 tensor<fp16, [1, 20, 64, 1500]> var_334_cast_fp16 = reshape(shape = var_333, x = value_3_cast_fp16)[name = tensor<string, []>("op_334_cast_fp16")];
247 tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
248 tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
249 tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_334_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
250 tensor<int32, [4]> var_337 = const()[name = tensor<string, []>("op_337"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
251 tensor<fp16, [1, 1280, 1, 1]> input_17_cast_fp16 = reshape(shape = var_337, x = attn_3_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
252 tensor<string, []> pretrained_out_15_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_15_pad_type_0"), val = tensor<string, []>("valid")];
253 tensor<int32, [2]> pretrained_out_15_strides_0 = const()[name = tensor<string, []>("pretrained_out_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
254 tensor<int32, [4]> pretrained_out_15_pad_0 = const()[name = tensor<string, []>("pretrained_out_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
255 tensor<int32, [2]> pretrained_out_15_dilations_0 = const()[name = tensor<string, []>("pretrained_out_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
256 tensor<int32, []> pretrained_out_15_groups_0 = const()[name = tensor<string, []>("pretrained_out_15_groups_0"), val = tensor<int32, []>(1)];
257 tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140262976))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141082240))), name = tensor<string, []>("layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
258 tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141082368)))];
259 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_15_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_15_dilations_0, groups = pretrained_out_15_groups_0, pad = pretrained_out_15_pad_0, pad_type = pretrained_out_15_pad_type_0, strides = pretrained_out_15_strides_0, weight = layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor<string, []>("pretrained_out_15_cast_fp16")];
260 tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
261 tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
262 tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
263 tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
264 tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
265 tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141084992)))];
266 tensor<fp16, [1, 16, 1, 1]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_0_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
267 tensor<string, []> lora_out_15_pad_type_0 = const()[name = tensor<string, []>("lora_out_15_pad_type_0"), val = tensor<string, []>("valid")];
268 tensor<int32, [2]> lora_out_15_strides_0 = const()[name = tensor<string, []>("lora_out_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
269 tensor<int32, [4]> lora_out_15_pad_0 = const()[name = tensor<string, []>("lora_out_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
270 tensor<int32, [2]> lora_out_15_dilations_0 = const()[name = tensor<string, []>("lora_out_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
271 tensor<int32, []> lora_out_15_groups_0 = const()[name = tensor<string, []>("lora_out_15_groups_0"), val = tensor<int32, []>(1)];
272 tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141126016)))];
273 tensor<fp16, [1, 1280, 1, 1]> lora_out_15_cast_fp16 = conv(dilations = lora_out_15_dilations_0, groups = lora_out_15_groups_0, pad = lora_out_15_pad_0, pad_type = lora_out_15_pad_type_0, strides = lora_out_15_strides_0, weight = layers_0_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("lora_out_15_cast_fp16")];
274 tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = add(x = pretrained_out_15_cast_fp16, y = lora_out_15_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
275 tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
276 tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
277 tensor<fp16, []> var_371_to_fp16 = const()[name = tensor<string, []>("op_371_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
278 tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_371_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
279 tensor<fp16, [1280]> input_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141167040)))];
280 tensor<fp16, [1280]> input_21_beta_0_to_fp16 = const()[name = tensor<string, []>("input_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141169664)))];
281 tensor<fp16, []> input_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
282 tensor<fp16, [1, 1280, 1, 1]> input_21_cast_fp16 = batch_norm(beta = input_21_beta_0_to_fp16, epsilon = input_21_epsilon_0_to_fp16, gamma = input_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
283 tensor<string, []> pretrained_out_17_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_17_pad_type_0"), val = tensor<string, []>("valid")];
284 tensor<int32, [2]> pretrained_out_17_strides_0 = const()[name = tensor<string, []>("pretrained_out_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
285 tensor<int32, [4]> pretrained_out_17_pad_0 = const()[name = tensor<string, []>("pretrained_out_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
286 tensor<int32, [2]> pretrained_out_17_dilations_0 = const()[name = tensor<string, []>("pretrained_out_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
287 tensor<int32, []> pretrained_out_17_groups_0 = const()[name = tensor<string, []>("pretrained_out_17_groups_0"), val = tensor<int32, []>(1)];
288 tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141172288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144449152))), name = tensor<string, []>("layers_0_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
289 tensor<fp16, [5120]> layers_0_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144449280)))];
290 tensor<fp16, [1, 5120, 1, 1]> pretrained_out_17_cast_fp16 = conv(bias = layers_0_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_17_dilations_0, groups = pretrained_out_17_groups_0, pad = pretrained_out_17_pad_0, pad_type = pretrained_out_17_pad_type_0, strides = pretrained_out_17_strides_0, weight = layers_0_fc1_pretrained_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor<string, []>("pretrained_out_17_cast_fp16")];
291 tensor<string, []> input_23_pad_type_0 = const()[name = tensor<string, []>("input_23_pad_type_0"), val = tensor<string, []>("valid")];
292 tensor<int32, [2]> input_23_strides_0 = const()[name = tensor<string, []>("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
293 tensor<int32, [4]> input_23_pad_0 = const()[name = tensor<string, []>("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
294 tensor<int32, [2]> input_23_dilations_0 = const()[name = tensor<string, []>("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
295 tensor<int32, []> input_23_groups_0 = const()[name = tensor<string, []>("input_23_groups_0"), val = tensor<int32, []>(1)];
296 tensor<fp16, [16, 1280, 1, 1]> layers_0_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144459584)))];
297 tensor<fp16, [1, 16, 1, 1]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_0_fc1_loraA_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
298 tensor<string, []> lora_out_17_pad_type_0 = const()[name = tensor<string, []>("lora_out_17_pad_type_0"), val = tensor<string, []>("valid")];
299 tensor<int32, [2]> lora_out_17_strides_0 = const()[name = tensor<string, []>("lora_out_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
300 tensor<int32, [4]> lora_out_17_pad_0 = const()[name = tensor<string, []>("lora_out_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
301 tensor<int32, [2]> lora_out_17_dilations_0 = const()[name = tensor<string, []>("lora_out_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
302 tensor<int32, []> lora_out_17_groups_0 = const()[name = tensor<string, []>("lora_out_17_groups_0"), val = tensor<int32, []>(1)];
303 tensor<fp16, [5120, 16, 1, 1]> layers_0_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144500608)))];
304 tensor<fp16, [1, 5120, 1, 1]> lora_out_17_cast_fp16 = conv(dilations = lora_out_17_dilations_0, groups = lora_out_17_groups_0, pad = lora_out_17_pad_0, pad_type = lora_out_17_pad_type_0, strides = lora_out_17_strides_0, weight = layers_0_fc1_loraB_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("lora_out_17_cast_fp16")];
305 tensor<fp16, [1, 5120, 1, 1]> input_25_cast_fp16 = add(x = pretrained_out_17_cast_fp16, y = lora_out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
306 tensor<string, []> input_27_mode_0 = const()[name = tensor<string, []>("input_27_mode_0"), val = tensor<string, []>("EXACT")];
307 tensor<fp16, [1, 5120, 1, 1]> input_27_cast_fp16 = gelu(mode = input_27_mode_0, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
308 tensor<string, []> pretrained_out_19_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_19_pad_type_0"), val = tensor<string, []>("valid")];
309 tensor<int32, [2]> pretrained_out_19_strides_0 = const()[name = tensor<string, []>("pretrained_out_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
310 tensor<int32, [4]> pretrained_out_19_pad_0 = const()[name = tensor<string, []>("pretrained_out_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
311 tensor<int32, [2]> pretrained_out_19_dilations_0 = const()[name = tensor<string, []>("pretrained_out_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
312 tensor<int32, []> pretrained_out_19_groups_0 = const()[name = tensor<string, []>("pretrained_out_19_groups_0"), val = tensor<int32, []>(1)];
313 tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144664512))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147941376))), name = tensor<string, []>("layers_0_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
314 tensor<fp16, [1280]> layers_0_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147941504)))];
315 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_19_cast_fp16 = conv(bias = layers_0_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_19_dilations_0, groups = pretrained_out_19_groups_0, pad = pretrained_out_19_pad_0, pad_type = pretrained_out_19_pad_type_0, strides = pretrained_out_19_strides_0, weight = layers_0_fc2_pretrained_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor<string, []>("pretrained_out_19_cast_fp16")];
316 tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")];
317 tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
318 tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
319 tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
320 tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
321 tensor<fp16, [16, 5120, 1, 1]> layers_0_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147944128)))];
322 tensor<fp16, [1, 16, 1, 1]> input_29_cast_fp16 = conv(dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_0_fc2_loraA_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
323 tensor<string, []> lora_out_19_pad_type_0 = const()[name = tensor<string, []>("lora_out_19_pad_type_0"), val = tensor<string, []>("valid")];
324 tensor<int32, [2]> lora_out_19_strides_0 = const()[name = tensor<string, []>("lora_out_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
325 tensor<int32, [4]> lora_out_19_pad_0 = const()[name = tensor<string, []>("lora_out_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
326 tensor<int32, [2]> lora_out_19_dilations_0 = const()[name = tensor<string, []>("lora_out_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
327 tensor<int32, []> lora_out_19_groups_0 = const()[name = tensor<string, []>("lora_out_19_groups_0"), val = tensor<int32, []>(1)];
328 tensor<fp16, [1280, 16, 1, 1]> layers_0_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148108032)))];
329 tensor<fp16, [1, 1280, 1, 1]> lora_out_19_cast_fp16 = conv(dilations = lora_out_19_dilations_0, groups = lora_out_19_groups_0, pad = lora_out_19_pad_0, pad_type = lora_out_19_pad_type_0, strides = lora_out_19_strides_0, weight = layers_0_fc2_loraB_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("lora_out_19_cast_fp16")];
330 tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = add(x = pretrained_out_19_cast_fp16, y = lora_out_19_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
331 tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
332 tensor<int32, []> var_438 = const()[name = tensor<string, []>("op_438"), val = tensor<int32, []>(3)];
333 tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
334 tensor<fp16, []> var_464_to_fp16 = const()[name = tensor<string, []>("op_464_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
335 tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_464_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
336 tensor<fp16, [1280]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148149056)))];
337 tensor<fp16, [1280]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148151680)))];
338 tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
339 tensor<fp16, [1, 1280, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
340 tensor<string, []> pretrained_out_21_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_21_pad_type_0"), val = tensor<string, []>("valid")];
341 tensor<int32, [2]> pretrained_out_21_strides_0 = const()[name = tensor<string, []>("pretrained_out_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
342 tensor<int32, [4]> pretrained_out_21_pad_0 = const()[name = tensor<string, []>("pretrained_out_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
343 tensor<int32, [2]> pretrained_out_21_dilations_0 = const()[name = tensor<string, []>("pretrained_out_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
344 tensor<int32, []> pretrained_out_21_groups_0 = const()[name = tensor<string, []>("pretrained_out_21_groups_0"), val = tensor<int32, []>(1)];
345 tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148154304))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148973568))), name = tensor<string, []>("layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
346 tensor<fp16, [1280]> layers_1_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148973696)))];
347 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_21_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_21_dilations_0, groups = pretrained_out_21_groups_0, pad = pretrained_out_21_pad_0, pad_type = pretrained_out_21_pad_type_0, strides = pretrained_out_21_strides_0, weight = layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("pretrained_out_21_cast_fp16")];
348 tensor<string, []> input_31_pad_type_0 = const()[name = tensor<string, []>("input_31_pad_type_0"), val = tensor<string, []>("valid")];
349 tensor<int32, [2]> input_31_strides_0 = const()[name = tensor<string, []>("input_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
350 tensor<int32, [4]> input_31_pad_0 = const()[name = tensor<string, []>("input_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
351 tensor<int32, [2]> input_31_dilations_0 = const()[name = tensor<string, []>("input_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
352 tensor<int32, []> input_31_groups_0 = const()[name = tensor<string, []>("input_31_groups_0"), val = tensor<int32, []>(1)];
353 tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148976320)))];
354 tensor<fp16, [1, 16, 1, 1]> input_31_cast_fp16 = conv(dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = layers_1_self_attn_q_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
355 tensor<string, []> lora_out_21_pad_type_0 = const()[name = tensor<string, []>("lora_out_21_pad_type_0"), val = tensor<string, []>("valid")];
356 tensor<int32, [2]> lora_out_21_strides_0 = const()[name = tensor<string, []>("lora_out_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
357 tensor<int32, [4]> lora_out_21_pad_0 = const()[name = tensor<string, []>("lora_out_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
358 tensor<int32, [2]> lora_out_21_dilations_0 = const()[name = tensor<string, []>("lora_out_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
359 tensor<int32, []> lora_out_21_groups_0 = const()[name = tensor<string, []>("lora_out_21_groups_0"), val = tensor<int32, []>(1)];
360 tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149017344)))];
361 tensor<fp16, [1, 1280, 1, 1]> lora_out_21_cast_fp16 = conv(dilations = lora_out_21_dilations_0, groups = lora_out_21_groups_0, pad = lora_out_21_pad_0, pad_type = lora_out_21_pad_type_0, strides = lora_out_21_strides_0, weight = layers_1_self_attn_q_proj_loraB_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("lora_out_21_cast_fp16")];
362 tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = add(x = pretrained_out_21_cast_fp16, y = lora_out_21_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
363 tensor<string, []> pretrained_out_23_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_23_pad_type_0"), val = tensor<string, []>("valid")];
364 tensor<int32, [2]> pretrained_out_23_strides_0 = const()[name = tensor<string, []>("pretrained_out_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
365 tensor<int32, [4]> pretrained_out_23_pad_0 = const()[name = tensor<string, []>("pretrained_out_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
366 tensor<int32, [2]> pretrained_out_23_dilations_0 = const()[name = tensor<string, []>("pretrained_out_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
367 tensor<int32, []> pretrained_out_23_groups_0 = const()[name = tensor<string, []>("pretrained_out_23_groups_0"), val = tensor<int32, []>(1)];
368 tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149058368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149877632))), name = tensor<string, []>("layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
369 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_23_cast_fp16 = conv(dilations = pretrained_out_23_dilations_0, groups = pretrained_out_23_groups_0, pad = pretrained_out_23_pad_0, pad_type = pretrained_out_23_pad_type_0, strides = pretrained_out_23_strides_0, weight = layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("pretrained_out_23_cast_fp16")];
370 tensor<string, []> input_33_pad_type_0 = const()[name = tensor<string, []>("input_33_pad_type_0"), val = tensor<string, []>("valid")];
371 tensor<int32, [2]> input_33_strides_0 = const()[name = tensor<string, []>("input_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
372 tensor<int32, [4]> input_33_pad_0 = const()[name = tensor<string, []>("input_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
373 tensor<int32, [2]> input_33_dilations_0 = const()[name = tensor<string, []>("input_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
374 tensor<int32, []> input_33_groups_0 = const()[name = tensor<string, []>("input_33_groups_0"), val = tensor<int32, []>(1)];
375 tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149877760)))];
376 tensor<fp16, [1, 16, 1, 1]> input_33_cast_fp16 = conv(dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = layers_1_self_attn_k_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
377 tensor<string, []> lora_out_23_pad_type_0 = const()[name = tensor<string, []>("lora_out_23_pad_type_0"), val = tensor<string, []>("valid")];
378 tensor<int32, [2]> lora_out_23_strides_0 = const()[name = tensor<string, []>("lora_out_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
379 tensor<int32, [4]> lora_out_23_pad_0 = const()[name = tensor<string, []>("lora_out_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
380 tensor<int32, [2]> lora_out_23_dilations_0 = const()[name = tensor<string, []>("lora_out_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
381 tensor<int32, []> lora_out_23_groups_0 = const()[name = tensor<string, []>("lora_out_23_groups_0"), val = tensor<int32, []>(1)];
382 tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149918784)))];
383 tensor<fp16, [1, 1280, 1, 1]> lora_out_23_cast_fp16 = conv(dilations = lora_out_23_dilations_0, groups = lora_out_23_groups_0, pad = lora_out_23_pad_0, pad_type = lora_out_23_pad_type_0, strides = lora_out_23_strides_0, weight = layers_1_self_attn_k_proj_loraB_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("lora_out_23_cast_fp16")];
384 tensor<fp16, [1, 1280, 1, 1]> current_key_3_cast_fp16 = add(x = pretrained_out_23_cast_fp16, y = lora_out_23_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")];
385 tensor<string, []> pretrained_out_25_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_25_pad_type_0"), val = tensor<string, []>("valid")];
386 tensor<int32, [2]> pretrained_out_25_strides_0 = const()[name = tensor<string, []>("pretrained_out_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
387 tensor<int32, [4]> pretrained_out_25_pad_0 = const()[name = tensor<string, []>("pretrained_out_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
388 tensor<int32, [2]> pretrained_out_25_dilations_0 = const()[name = tensor<string, []>("pretrained_out_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
389 tensor<int32, []> pretrained_out_25_groups_0 = const()[name = tensor<string, []>("pretrained_out_25_groups_0"), val = tensor<int32, []>(1)];
390 tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149959808))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150779072))), name = tensor<string, []>("layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
391 tensor<fp16, [1280]> layers_1_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150779200)))];
392 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_25_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_25_dilations_0, groups = pretrained_out_25_groups_0, pad = pretrained_out_25_pad_0, pad_type = pretrained_out_25_pad_type_0, strides = pretrained_out_25_strides_0, weight = layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("pretrained_out_25_cast_fp16")];
393 tensor<string, []> input_35_pad_type_0 = const()[name = tensor<string, []>("input_35_pad_type_0"), val = tensor<string, []>("valid")];
394 tensor<int32, [2]> input_35_strides_0 = const()[name = tensor<string, []>("input_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
395 tensor<int32, [4]> input_35_pad_0 = const()[name = tensor<string, []>("input_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
396 tensor<int32, [2]> input_35_dilations_0 = const()[name = tensor<string, []>("input_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
397 tensor<int32, []> input_35_groups_0 = const()[name = tensor<string, []>("input_35_groups_0"), val = tensor<int32, []>(1)];
398 tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150781824)))];
399 tensor<fp16, [1, 16, 1, 1]> input_35_cast_fp16 = conv(dilations = input_35_dilations_0, groups = input_35_groups_0, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = input_35_strides_0, weight = layers_1_self_attn_v_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
400 tensor<string, []> lora_out_25_pad_type_0 = const()[name = tensor<string, []>("lora_out_25_pad_type_0"), val = tensor<string, []>("valid")];
401 tensor<int32, [2]> lora_out_25_strides_0 = const()[name = tensor<string, []>("lora_out_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
402 tensor<int32, [4]> lora_out_25_pad_0 = const()[name = tensor<string, []>("lora_out_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
403 tensor<int32, [2]> lora_out_25_dilations_0 = const()[name = tensor<string, []>("lora_out_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
404 tensor<int32, []> lora_out_25_groups_0 = const()[name = tensor<string, []>("lora_out_25_groups_0"), val = tensor<int32, []>(1)];
405 tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150822848)))];
406 tensor<fp16, [1, 1280, 1, 1]> lora_out_25_cast_fp16 = conv(dilations = lora_out_25_dilations_0, groups = lora_out_25_groups_0, pad = lora_out_25_pad_0, pad_type = lora_out_25_pad_type_0, strides = lora_out_25_strides_0, weight = layers_1_self_attn_v_proj_loraB_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("lora_out_25_cast_fp16")];
407 tensor<fp16, [1, 1280, 1, 1]> current_value_3_cast_fp16 = add(x = pretrained_out_25_cast_fp16, y = lora_out_25_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")];
408 tensor<fp16, [1, 1280, 1, 448]> var_550_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_550_cast_fp16")];
409 tensor<fp16, [1, 1280, 1, 448]> var_552_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_177_cast_fp16)[name = tensor<string, []>("op_552_cast_fp16")];
410 tensor<fp16, [1, 1280, 1, 448]> key_5_cast_fp16 = add(x = var_550_cast_fp16, y = var_552_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
411 tensor<fp16, [1, 1280, 1, 448]> var_554_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_554_cast_fp16")];
412 tensor<fp16, [1, 1280, 1, 448]> var_556_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_177_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")];
413 tensor<fp16, [1, 1280, 1, 448]> value_5_cast_fp16 = add(x = var_554_cast_fp16, y = var_556_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
414 tensor<int32, [4]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [4]>([1, 20, 64, -1])];
415 tensor<fp16, [1, 20, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_559, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
416 tensor<fp16, []> var_561_to_fp16 = const()[name = tensor<string, []>("op_561_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
417 tensor<fp16, [1, 20, 64, 1]> var_562_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_561_to_fp16)[name = tensor<string, []>("op_562_cast_fp16")];
418 tensor<int32, [4]> var_563 = const()[name = tensor<string, []>("op_563"), val = tensor<int32, [4]>([1, 20, 64, -1])];
419 tensor<fp16, [1, 20, 64, 448]> var_564_cast_fp16 = reshape(shape = var_563, x = key_5_cast_fp16)[name = tensor<string, []>("op_564_cast_fp16")];
420 tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
421 tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
422 tensor<fp16, [1, 20, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_562_cast_fp16, y = var_564_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
423 tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
424 tensor<fp16, [1, 20, 1, 448]> var_572_cast_fp16 = softmax(axis = var_438, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_572_cast_fp16")];
425 tensor<int32, [4]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [4]>([1, 20, 64, -1])];
426 tensor<fp16, [1, 20, 64, 448]> var_574_cast_fp16 = reshape(shape = var_573, x = value_5_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
427 tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
428 tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
429 tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_574_cast_fp16, y = var_572_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
430 tensor<int32, [4]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
431 tensor<fp16, [1, 1280, 1, 1]> input_37_cast_fp16 = reshape(shape = var_577, x = attn_5_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
432 tensor<string, []> pretrained_out_27_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_27_pad_type_0"), val = tensor<string, []>("valid")];
433 tensor<int32, [2]> pretrained_out_27_strides_0 = const()[name = tensor<string, []>("pretrained_out_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
434 tensor<int32, [4]> pretrained_out_27_pad_0 = const()[name = tensor<string, []>("pretrained_out_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
435 tensor<int32, [2]> pretrained_out_27_dilations_0 = const()[name = tensor<string, []>("pretrained_out_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
436 tensor<int32, []> pretrained_out_27_groups_0 = const()[name = tensor<string, []>("pretrained_out_27_groups_0"), val = tensor<int32, []>(1)];
437 tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150863872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151683136))), name = tensor<string, []>("layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
438 tensor<fp16, [1280]> layers_1_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151683264)))];
439 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_27_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_27_dilations_0, groups = pretrained_out_27_groups_0, pad = pretrained_out_27_pad_0, pad_type = pretrained_out_27_pad_type_0, strides = pretrained_out_27_strides_0, weight = layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor<string, []>("pretrained_out_27_cast_fp16")];
440 tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("valid")];
441 tensor<int32, [2]> input_39_strides_0 = const()[name = tensor<string, []>("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
442 tensor<int32, [4]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
443 tensor<int32, [2]> input_39_dilations_0 = const()[name = tensor<string, []>("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
444 tensor<int32, []> input_39_groups_0 = const()[name = tensor<string, []>("input_39_groups_0"), val = tensor<int32, []>(1)];
445 tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151685888)))];
446 tensor<fp16, [1, 16, 1, 1]> input_39_cast_fp16 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_1_self_attn_o_proj_loraA_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
447 tensor<string, []> lora_out_27_pad_type_0 = const()[name = tensor<string, []>("lora_out_27_pad_type_0"), val = tensor<string, []>("valid")];
448 tensor<int32, [2]> lora_out_27_strides_0 = const()[name = tensor<string, []>("lora_out_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
449 tensor<int32, [4]> lora_out_27_pad_0 = const()[name = tensor<string, []>("lora_out_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
450 tensor<int32, [2]> lora_out_27_dilations_0 = const()[name = tensor<string, []>("lora_out_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
451 tensor<int32, []> lora_out_27_groups_0 = const()[name = tensor<string, []>("lora_out_27_groups_0"), val = tensor<int32, []>(1)];
452 tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151726912)))];
453 tensor<fp16, [1, 1280, 1, 1]> lora_out_27_cast_fp16 = conv(dilations = lora_out_27_dilations_0, groups = lora_out_27_groups_0, pad = lora_out_27_pad_0, pad_type = lora_out_27_pad_type_0, strides = lora_out_27_strides_0, weight = layers_1_self_attn_o_proj_loraB_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("lora_out_27_cast_fp16")];
454 tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = add(x = pretrained_out_27_cast_fp16, y = lora_out_27_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
455 tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
456 tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
457 tensor<fp16, []> var_615_to_fp16 = const()[name = tensor<string, []>("op_615_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
458 tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_615_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
459 tensor<fp16, [1280]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151767936)))];
460 tensor<fp16, [1280]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151770560)))];
461 tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
462 tensor<fp16, [1, 1280, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
463 tensor<string, []> pretrained_out_29_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_29_pad_type_0"), val = tensor<string, []>("valid")];
464 tensor<int32, [2]> pretrained_out_29_strides_0 = const()[name = tensor<string, []>("pretrained_out_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
465 tensor<int32, [4]> pretrained_out_29_pad_0 = const()[name = tensor<string, []>("pretrained_out_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
466 tensor<int32, [2]> pretrained_out_29_dilations_0 = const()[name = tensor<string, []>("pretrained_out_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
467 tensor<int32, []> pretrained_out_29_groups_0 = const()[name = tensor<string, []>("pretrained_out_29_groups_0"), val = tensor<int32, []>(1)];
468 tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151773184))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152592448))), name = tensor<string, []>("layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
469 tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152592576)))];
470 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_29_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_29_dilations_0, groups = pretrained_out_29_groups_0, pad = pretrained_out_29_pad_0, pad_type = pretrained_out_29_pad_type_0, strides = pretrained_out_29_strides_0, weight = layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor<string, []>("pretrained_out_29_cast_fp16")];
471 tensor<string, []> input_41_pad_type_0 = const()[name = tensor<string, []>("input_41_pad_type_0"), val = tensor<string, []>("valid")];
472 tensor<int32, [2]> input_41_strides_0 = const()[name = tensor<string, []>("input_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
473 tensor<int32, [4]> input_41_pad_0 = const()[name = tensor<string, []>("input_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
474 tensor<int32, [2]> input_41_dilations_0 = const()[name = tensor<string, []>("input_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
475 tensor<int32, []> input_41_groups_0 = const()[name = tensor<string, []>("input_41_groups_0"), val = tensor<int32, []>(1)];
476 tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152595200)))];
477 tensor<fp16, [1, 16, 1, 1]> input_41_cast_fp16 = conv(dilations = input_41_dilations_0, groups = input_41_groups_0, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = input_41_strides_0, weight = layers_1_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
478 tensor<string, []> lora_out_29_pad_type_0 = const()[name = tensor<string, []>("lora_out_29_pad_type_0"), val = tensor<string, []>("valid")];
479 tensor<int32, [2]> lora_out_29_strides_0 = const()[name = tensor<string, []>("lora_out_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
480 tensor<int32, [4]> lora_out_29_pad_0 = const()[name = tensor<string, []>("lora_out_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
481 tensor<int32, [2]> lora_out_29_dilations_0 = const()[name = tensor<string, []>("lora_out_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
482 tensor<int32, []> lora_out_29_groups_0 = const()[name = tensor<string, []>("lora_out_29_groups_0"), val = tensor<int32, []>(1)];
483 tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152636224)))];
484 tensor<fp16, [1, 1280, 1, 1]> lora_out_29_cast_fp16 = conv(dilations = lora_out_29_dilations_0, groups = lora_out_29_groups_0, pad = lora_out_29_pad_0, pad_type = lora_out_29_pad_type_0, strides = lora_out_29_strides_0, weight = layers_1_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("lora_out_29_cast_fp16")];
485 tensor<fp16, [1, 1280, 1, 1]> query_7_cast_fp16 = add(x = pretrained_out_29_cast_fp16, y = lora_out_29_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
486 tensor<string, []> pretrained_out_31_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_31_pad_type_0"), val = tensor<string, []>("valid")];
487 tensor<int32, [2]> pretrained_out_31_strides_0 = const()[name = tensor<string, []>("pretrained_out_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
488 tensor<int32, [4]> pretrained_out_31_pad_0 = const()[name = tensor<string, []>("pretrained_out_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
489 tensor<int32, [2]> pretrained_out_31_dilations_0 = const()[name = tensor<string, []>("pretrained_out_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
490 tensor<int32, []> pretrained_out_31_groups_0 = const()[name = tensor<string, []>("pretrained_out_31_groups_0"), val = tensor<int32, []>(1)];
491 tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152677248))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153496512))), name = tensor<string, []>("layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
492 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_31_cast_fp16 = conv(dilations = pretrained_out_31_dilations_0, groups = pretrained_out_31_groups_0, pad = pretrained_out_31_pad_0, pad_type = pretrained_out_31_pad_type_0, strides = pretrained_out_31_strides_0, weight = layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_31_cast_fp16")];
493 tensor<string, []> input_43_pad_type_0 = const()[name = tensor<string, []>("input_43_pad_type_0"), val = tensor<string, []>("valid")];
494 tensor<int32, [2]> input_43_strides_0 = const()[name = tensor<string, []>("input_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
495 tensor<int32, [4]> input_43_pad_0 = const()[name = tensor<string, []>("input_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
496 tensor<int32, [2]> input_43_dilations_0 = const()[name = tensor<string, []>("input_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
497 tensor<int32, []> input_43_groups_0 = const()[name = tensor<string, []>("input_43_groups_0"), val = tensor<int32, []>(1)];
498 tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153496640)))];
499 tensor<fp16, [1, 16, 1, 1500]> input_43_cast_fp16 = conv(dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_1_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_43_cast_fp16")];
500 tensor<string, []> lora_out_31_pad_type_0 = const()[name = tensor<string, []>("lora_out_31_pad_type_0"), val = tensor<string, []>("valid")];
501 tensor<int32, [2]> lora_out_31_strides_0 = const()[name = tensor<string, []>("lora_out_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
502 tensor<int32, [4]> lora_out_31_pad_0 = const()[name = tensor<string, []>("lora_out_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
503 tensor<int32, [2]> lora_out_31_dilations_0 = const()[name = tensor<string, []>("lora_out_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
504 tensor<int32, []> lora_out_31_groups_0 = const()[name = tensor<string, []>("lora_out_31_groups_0"), val = tensor<int32, []>(1)];
505 tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153537664)))];
506 tensor<fp16, [1, 1280, 1, 1500]> lora_out_31_cast_fp16 = conv(dilations = lora_out_31_dilations_0, groups = lora_out_31_groups_0, pad = lora_out_31_pad_0, pad_type = lora_out_31_pad_type_0, strides = lora_out_31_strides_0, weight = layers_1_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("lora_out_31_cast_fp16")];
507 tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = add(x = pretrained_out_31_cast_fp16, y = lora_out_31_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
508 tensor<string, []> pretrained_out_33_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_33_pad_type_0"), val = tensor<string, []>("valid")];
509 tensor<int32, [2]> pretrained_out_33_strides_0 = const()[name = tensor<string, []>("pretrained_out_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
510 tensor<int32, [4]> pretrained_out_33_pad_0 = const()[name = tensor<string, []>("pretrained_out_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
511 tensor<int32, [2]> pretrained_out_33_dilations_0 = const()[name = tensor<string, []>("pretrained_out_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
512 tensor<int32, []> pretrained_out_33_groups_0 = const()[name = tensor<string, []>("pretrained_out_33_groups_0"), val = tensor<int32, []>(1)];
513 tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153578688))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154397952))), name = tensor<string, []>("layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
514 tensor<fp16, [1280]> layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154398080)))];
515 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_33_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_33_dilations_0, groups = pretrained_out_33_groups_0, pad = pretrained_out_33_pad_0, pad_type = pretrained_out_33_pad_type_0, strides = pretrained_out_33_strides_0, weight = layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_33_cast_fp16")];
516 tensor<string, []> input_45_pad_type_0 = const()[name = tensor<string, []>("input_45_pad_type_0"), val = tensor<string, []>("valid")];
517 tensor<int32, [2]> input_45_strides_0 = const()[name = tensor<string, []>("input_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
518 tensor<int32, [4]> input_45_pad_0 = const()[name = tensor<string, []>("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
519 tensor<int32, [2]> input_45_dilations_0 = const()[name = tensor<string, []>("input_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
520 tensor<int32, []> input_45_groups_0 = const()[name = tensor<string, []>("input_45_groups_0"), val = tensor<int32, []>(1)];
521 tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154400704)))];
522 tensor<fp16, [1, 16, 1, 1500]> input_45_cast_fp16 = conv(dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_1_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_45_cast_fp16")];
523 tensor<string, []> lora_out_33_pad_type_0 = const()[name = tensor<string, []>("lora_out_33_pad_type_0"), val = tensor<string, []>("valid")];
524 tensor<int32, [2]> lora_out_33_strides_0 = const()[name = tensor<string, []>("lora_out_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
525 tensor<int32, [4]> lora_out_33_pad_0 = const()[name = tensor<string, []>("lora_out_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
526 tensor<int32, [2]> lora_out_33_dilations_0 = const()[name = tensor<string, []>("lora_out_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
527 tensor<int32, []> lora_out_33_groups_0 = const()[name = tensor<string, []>("lora_out_33_groups_0"), val = tensor<int32, []>(1)];
528 tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154441728)))];
529 tensor<fp16, [1, 1280, 1, 1500]> lora_out_33_cast_fp16 = conv(dilations = lora_out_33_dilations_0, groups = lora_out_33_groups_0, pad = lora_out_33_pad_0, pad_type = lora_out_33_pad_type_0, strides = lora_out_33_strides_0, weight = layers_1_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("lora_out_33_cast_fp16")];
530 tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = add(x = pretrained_out_33_cast_fp16, y = lora_out_33_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")];
531 tensor<int32, [4]> var_698 = const()[name = tensor<string, []>("op_698"), val = tensor<int32, [4]>([1, 20, 64, -1])];
532 tensor<fp16, [1, 20, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_698, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
533 tensor<fp16, []> var_700_to_fp16 = const()[name = tensor<string, []>("op_700_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
534 tensor<fp16, [1, 20, 64, 1]> var_701_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_700_to_fp16)[name = tensor<string, []>("op_701_cast_fp16")];
535 tensor<int32, [4]> var_702 = const()[name = tensor<string, []>("op_702"), val = tensor<int32, [4]>([1, 20, 64, -1])];
536 tensor<fp16, [1, 20, 64, 1500]> var_703_cast_fp16 = reshape(shape = var_702, x = key_7_cast_fp16)[name = tensor<string, []>("op_703_cast_fp16")];
537 tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
538 tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
539 tensor<fp16, [1, 20, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_701_cast_fp16, y = var_703_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
540 tensor<fp16, [1, 20, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_438, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
541 tensor<int32, [4]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [4]>([1, 20, 64, -1])];
542 tensor<fp16, [1, 20, 64, 1500]> var_708_cast_fp16 = reshape(shape = var_707, x = value_7_cast_fp16)[name = tensor<string, []>("op_708_cast_fp16")];
543 tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
544 tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
545 tensor<fp16, [1, 20, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_708_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
546 tensor<int32, [4]> var_711 = const()[name = tensor<string, []>("op_711"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
547 tensor<fp16, [1, 1280, 1, 1]> input_47_cast_fp16 = reshape(shape = var_711, x = attn_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
548 tensor<string, []> pretrained_out_35_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_35_pad_type_0"), val = tensor<string, []>("valid")];
549 tensor<int32, [2]> pretrained_out_35_strides_0 = const()[name = tensor<string, []>("pretrained_out_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
550 tensor<int32, [4]> pretrained_out_35_pad_0 = const()[name = tensor<string, []>("pretrained_out_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
551 tensor<int32, [2]> pretrained_out_35_dilations_0 = const()[name = tensor<string, []>("pretrained_out_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
552 tensor<int32, []> pretrained_out_35_groups_0 = const()[name = tensor<string, []>("pretrained_out_35_groups_0"), val = tensor<int32, []>(1)];
553 tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154482752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155302016))), name = tensor<string, []>("layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
554 tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155302144)))];
555 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_35_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_35_dilations_0, groups = pretrained_out_35_groups_0, pad = pretrained_out_35_pad_0, pad_type = pretrained_out_35_pad_type_0, strides = pretrained_out_35_strides_0, weight = layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor<string, []>("pretrained_out_35_cast_fp16")];
556 tensor<string, []> input_49_pad_type_0 = const()[name = tensor<string, []>("input_49_pad_type_0"), val = tensor<string, []>("valid")];
557 tensor<int32, [2]> input_49_strides_0 = const()[name = tensor<string, []>("input_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
558 tensor<int32, [4]> input_49_pad_0 = const()[name = tensor<string, []>("input_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
559 tensor<int32, [2]> input_49_dilations_0 = const()[name = tensor<string, []>("input_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
560 tensor<int32, []> input_49_groups_0 = const()[name = tensor<string, []>("input_49_groups_0"), val = tensor<int32, []>(1)];
561 tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155304768)))];
562 tensor<fp16, [1, 16, 1, 1]> input_49_cast_fp16 = conv(dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = layers_1_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
563 tensor<string, []> lora_out_35_pad_type_0 = const()[name = tensor<string, []>("lora_out_35_pad_type_0"), val = tensor<string, []>("valid")];
564 tensor<int32, [2]> lora_out_35_strides_0 = const()[name = tensor<string, []>("lora_out_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
565 tensor<int32, [4]> lora_out_35_pad_0 = const()[name = tensor<string, []>("lora_out_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
566 tensor<int32, [2]> lora_out_35_dilations_0 = const()[name = tensor<string, []>("lora_out_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
567 tensor<int32, []> lora_out_35_groups_0 = const()[name = tensor<string, []>("lora_out_35_groups_0"), val = tensor<int32, []>(1)];
568 tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155345792)))];
569 tensor<fp16, [1, 1280, 1, 1]> lora_out_35_cast_fp16 = conv(dilations = lora_out_35_dilations_0, groups = lora_out_35_groups_0, pad = lora_out_35_pad_0, pad_type = lora_out_35_pad_type_0, strides = lora_out_35_strides_0, weight = layers_1_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("lora_out_35_cast_fp16")];
570 tensor<fp16, [1, 1280, 1, 1]> obj_25_cast_fp16 = add(x = pretrained_out_35_cast_fp16, y = lora_out_35_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
571 tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
572 tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
573 tensor<fp16, []> var_745_to_fp16 = const()[name = tensor<string, []>("op_745_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
574 tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_745_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
575 tensor<fp16, [1280]> input_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155386816)))];
576 tensor<fp16, [1280]> input_51_beta_0_to_fp16 = const()[name = tensor<string, []>("input_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155389440)))];
577 tensor<fp16, []> input_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
578 tensor<fp16, [1, 1280, 1, 1]> input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
579 tensor<string, []> pretrained_out_37_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_37_pad_type_0"), val = tensor<string, []>("valid")];
580 tensor<int32, [2]> pretrained_out_37_strides_0 = const()[name = tensor<string, []>("pretrained_out_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
581 tensor<int32, [4]> pretrained_out_37_pad_0 = const()[name = tensor<string, []>("pretrained_out_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
582 tensor<int32, [2]> pretrained_out_37_dilations_0 = const()[name = tensor<string, []>("pretrained_out_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
583 tensor<int32, []> pretrained_out_37_groups_0 = const()[name = tensor<string, []>("pretrained_out_37_groups_0"), val = tensor<int32, []>(1)];
584 tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155392064))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158668928))), name = tensor<string, []>("layers_1_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
585 tensor<fp16, [5120]> layers_1_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158669056)))];
586 tensor<fp16, [1, 5120, 1, 1]> pretrained_out_37_cast_fp16 = conv(bias = layers_1_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_37_dilations_0, groups = pretrained_out_37_groups_0, pad = pretrained_out_37_pad_0, pad_type = pretrained_out_37_pad_type_0, strides = pretrained_out_37_strides_0, weight = layers_1_fc1_pretrained_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor<string, []>("pretrained_out_37_cast_fp16")];
587 tensor<string, []> input_53_pad_type_0 = const()[name = tensor<string, []>("input_53_pad_type_0"), val = tensor<string, []>("valid")];
588 tensor<int32, [2]> input_53_strides_0 = const()[name = tensor<string, []>("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
589 tensor<int32, [4]> input_53_pad_0 = const()[name = tensor<string, []>("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
590 tensor<int32, [2]> input_53_dilations_0 = const()[name = tensor<string, []>("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
591 tensor<int32, []> input_53_groups_0 = const()[name = tensor<string, []>("input_53_groups_0"), val = tensor<int32, []>(1)];
592 tensor<fp16, [16, 1280, 1, 1]> layers_1_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158679360)))];
593 tensor<fp16, [1, 16, 1, 1]> input_53_cast_fp16 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_1_fc1_loraA_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
594 tensor<string, []> lora_out_37_pad_type_0 = const()[name = tensor<string, []>("lora_out_37_pad_type_0"), val = tensor<string, []>("valid")];
595 tensor<int32, [2]> lora_out_37_strides_0 = const()[name = tensor<string, []>("lora_out_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
596 tensor<int32, [4]> lora_out_37_pad_0 = const()[name = tensor<string, []>("lora_out_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
597 tensor<int32, [2]> lora_out_37_dilations_0 = const()[name = tensor<string, []>("lora_out_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
598 tensor<int32, []> lora_out_37_groups_0 = const()[name = tensor<string, []>("lora_out_37_groups_0"), val = tensor<int32, []>(1)];
599 tensor<fp16, [5120, 16, 1, 1]> layers_1_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158720384)))];
600 tensor<fp16, [1, 5120, 1, 1]> lora_out_37_cast_fp16 = conv(dilations = lora_out_37_dilations_0, groups = lora_out_37_groups_0, pad = lora_out_37_pad_0, pad_type = lora_out_37_pad_type_0, strides = lora_out_37_strides_0, weight = layers_1_fc1_loraB_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("lora_out_37_cast_fp16")];
601 tensor<fp16, [1, 5120, 1, 1]> input_55_cast_fp16 = add(x = pretrained_out_37_cast_fp16, y = lora_out_37_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
602 tensor<string, []> input_57_mode_0 = const()[name = tensor<string, []>("input_57_mode_0"), val = tensor<string, []>("EXACT")];
603 tensor<fp16, [1, 5120, 1, 1]> input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
604 tensor<string, []> pretrained_out_39_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_39_pad_type_0"), val = tensor<string, []>("valid")];
605 tensor<int32, [2]> pretrained_out_39_strides_0 = const()[name = tensor<string, []>("pretrained_out_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
606 tensor<int32, [4]> pretrained_out_39_pad_0 = const()[name = tensor<string, []>("pretrained_out_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
607 tensor<int32, [2]> pretrained_out_39_dilations_0 = const()[name = tensor<string, []>("pretrained_out_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
608 tensor<int32, []> pretrained_out_39_groups_0 = const()[name = tensor<string, []>("pretrained_out_39_groups_0"), val = tensor<int32, []>(1)];
609 tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158884288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162161152))), name = tensor<string, []>("layers_1_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
610 tensor<fp16, [1280]> layers_1_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162161280)))];
611 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_39_cast_fp16 = conv(bias = layers_1_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_39_dilations_0, groups = pretrained_out_39_groups_0, pad = pretrained_out_39_pad_0, pad_type = pretrained_out_39_pad_type_0, strides = pretrained_out_39_strides_0, weight = layers_1_fc2_pretrained_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor<string, []>("pretrained_out_39_cast_fp16")];
612 tensor<string, []> input_59_pad_type_0 = const()[name = tensor<string, []>("input_59_pad_type_0"), val = tensor<string, []>("valid")];
613 tensor<int32, [2]> input_59_strides_0 = const()[name = tensor<string, []>("input_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
614 tensor<int32, [4]> input_59_pad_0 = const()[name = tensor<string, []>("input_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
615 tensor<int32, [2]> input_59_dilations_0 = const()[name = tensor<string, []>("input_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
616 tensor<int32, []> input_59_groups_0 = const()[name = tensor<string, []>("input_59_groups_0"), val = tensor<int32, []>(1)];
617 tensor<fp16, [16, 5120, 1, 1]> layers_1_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162163904)))];
618 tensor<fp16, [1, 16, 1, 1]> input_59_cast_fp16 = conv(dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = layers_1_fc2_loraA_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
619 tensor<string, []> lora_out_39_pad_type_0 = const()[name = tensor<string, []>("lora_out_39_pad_type_0"), val = tensor<string, []>("valid")];
620 tensor<int32, [2]> lora_out_39_strides_0 = const()[name = tensor<string, []>("lora_out_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
621 tensor<int32, [4]> lora_out_39_pad_0 = const()[name = tensor<string, []>("lora_out_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
622 tensor<int32, [2]> lora_out_39_dilations_0 = const()[name = tensor<string, []>("lora_out_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
623 tensor<int32, []> lora_out_39_groups_0 = const()[name = tensor<string, []>("lora_out_39_groups_0"), val = tensor<int32, []>(1)];
624 tensor<fp16, [1280, 16, 1, 1]> layers_1_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162327808)))];
625 tensor<fp16, [1, 1280, 1, 1]> lora_out_39_cast_fp16 = conv(dilations = lora_out_39_dilations_0, groups = lora_out_39_groups_0, pad = lora_out_39_pad_0, pad_type = lora_out_39_pad_type_0, strides = lora_out_39_strides_0, weight = layers_1_fc2_loraB_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("lora_out_39_cast_fp16")];
626 tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = add(x = pretrained_out_39_cast_fp16, y = lora_out_39_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
627 tensor<fp16, [1, 1280, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
628 tensor<int32, []> var_812 = const()[name = tensor<string, []>("op_812"), val = tensor<int32, []>(3)];
629 tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
630 tensor<fp16, []> var_838_to_fp16 = const()[name = tensor<string, []>("op_838_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
631 tensor<fp16, [1, 1280, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_838_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
632 tensor<fp16, [1280]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162368832)))];
633 tensor<fp16, [1280]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162371456)))];
634 tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
635 tensor<fp16, [1, 1280, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
636 tensor<string, []> pretrained_out_41_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_41_pad_type_0"), val = tensor<string, []>("valid")];
637 tensor<int32, [2]> pretrained_out_41_strides_0 = const()[name = tensor<string, []>("pretrained_out_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
638 tensor<int32, [4]> pretrained_out_41_pad_0 = const()[name = tensor<string, []>("pretrained_out_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
639 tensor<int32, [2]> pretrained_out_41_dilations_0 = const()[name = tensor<string, []>("pretrained_out_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
640 tensor<int32, []> pretrained_out_41_groups_0 = const()[name = tensor<string, []>("pretrained_out_41_groups_0"), val = tensor<int32, []>(1)];
641 tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162374080))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163193344))), name = tensor<string, []>("layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
642 tensor<fp16, [1280]> layers_2_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163193472)))];
643 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_41_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_41_dilations_0, groups = pretrained_out_41_groups_0, pad = pretrained_out_41_pad_0, pad_type = pretrained_out_41_pad_type_0, strides = pretrained_out_41_strides_0, weight = layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("pretrained_out_41_cast_fp16")];
644 tensor<string, []> input_61_pad_type_0 = const()[name = tensor<string, []>("input_61_pad_type_0"), val = tensor<string, []>("valid")];
645 tensor<int32, [2]> input_61_strides_0 = const()[name = tensor<string, []>("input_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
646 tensor<int32, [4]> input_61_pad_0 = const()[name = tensor<string, []>("input_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
647 tensor<int32, [2]> input_61_dilations_0 = const()[name = tensor<string, []>("input_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
648 tensor<int32, []> input_61_groups_0 = const()[name = tensor<string, []>("input_61_groups_0"), val = tensor<int32, []>(1)];
649 tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163196096)))];
650 tensor<fp16, [1, 16, 1, 1]> input_61_cast_fp16 = conv(dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_2_self_attn_q_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
651 tensor<string, []> lora_out_41_pad_type_0 = const()[name = tensor<string, []>("lora_out_41_pad_type_0"), val = tensor<string, []>("valid")];
652 tensor<int32, [2]> lora_out_41_strides_0 = const()[name = tensor<string, []>("lora_out_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
653 tensor<int32, [4]> lora_out_41_pad_0 = const()[name = tensor<string, []>("lora_out_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
654 tensor<int32, [2]> lora_out_41_dilations_0 = const()[name = tensor<string, []>("lora_out_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
655 tensor<int32, []> lora_out_41_groups_0 = const()[name = tensor<string, []>("lora_out_41_groups_0"), val = tensor<int32, []>(1)];
656 tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163237120)))];
657 tensor<fp16, [1, 1280, 1, 1]> lora_out_41_cast_fp16 = conv(dilations = lora_out_41_dilations_0, groups = lora_out_41_groups_0, pad = lora_out_41_pad_0, pad_type = lora_out_41_pad_type_0, strides = lora_out_41_strides_0, weight = layers_2_self_attn_q_proj_loraB_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("lora_out_41_cast_fp16")];
658 tensor<fp16, [1, 1280, 1, 1]> query_9_cast_fp16 = add(x = pretrained_out_41_cast_fp16, y = lora_out_41_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
659 tensor<string, []> pretrained_out_43_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_43_pad_type_0"), val = tensor<string, []>("valid")];
660 tensor<int32, [2]> pretrained_out_43_strides_0 = const()[name = tensor<string, []>("pretrained_out_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
661 tensor<int32, [4]> pretrained_out_43_pad_0 = const()[name = tensor<string, []>("pretrained_out_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
662 tensor<int32, [2]> pretrained_out_43_dilations_0 = const()[name = tensor<string, []>("pretrained_out_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
663 tensor<int32, []> pretrained_out_43_groups_0 = const()[name = tensor<string, []>("pretrained_out_43_groups_0"), val = tensor<int32, []>(1)];
664 tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163278144))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164097408))), name = tensor<string, []>("layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
665 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_43_cast_fp16 = conv(dilations = pretrained_out_43_dilations_0, groups = pretrained_out_43_groups_0, pad = pretrained_out_43_pad_0, pad_type = pretrained_out_43_pad_type_0, strides = pretrained_out_43_strides_0, weight = layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("pretrained_out_43_cast_fp16")];
666 tensor<string, []> input_63_pad_type_0 = const()[name = tensor<string, []>("input_63_pad_type_0"), val = tensor<string, []>("valid")];
667 tensor<int32, [2]> input_63_strides_0 = const()[name = tensor<string, []>("input_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
668 tensor<int32, [4]> input_63_pad_0 = const()[name = tensor<string, []>("input_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
669 tensor<int32, [2]> input_63_dilations_0 = const()[name = tensor<string, []>("input_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
670 tensor<int32, []> input_63_groups_0 = const()[name = tensor<string, []>("input_63_groups_0"), val = tensor<int32, []>(1)];
671 tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164097536)))];
672 tensor<fp16, [1, 16, 1, 1]> input_63_cast_fp16 = conv(dilations = input_63_dilations_0, groups = input_63_groups_0, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = input_63_strides_0, weight = layers_2_self_attn_k_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
673 tensor<string, []> lora_out_43_pad_type_0 = const()[name = tensor<string, []>("lora_out_43_pad_type_0"), val = tensor<string, []>("valid")];
674 tensor<int32, [2]> lora_out_43_strides_0 = const()[name = tensor<string, []>("lora_out_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
675 tensor<int32, [4]> lora_out_43_pad_0 = const()[name = tensor<string, []>("lora_out_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
676 tensor<int32, [2]> lora_out_43_dilations_0 = const()[name = tensor<string, []>("lora_out_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
677 tensor<int32, []> lora_out_43_groups_0 = const()[name = tensor<string, []>("lora_out_43_groups_0"), val = tensor<int32, []>(1)];
678 tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164138560)))];
679 tensor<fp16, [1, 1280, 1, 1]> lora_out_43_cast_fp16 = conv(dilations = lora_out_43_dilations_0, groups = lora_out_43_groups_0, pad = lora_out_43_pad_0, pad_type = lora_out_43_pad_type_0, strides = lora_out_43_strides_0, weight = layers_2_self_attn_k_proj_loraB_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("lora_out_43_cast_fp16")];
680 tensor<fp16, [1, 1280, 1, 1]> current_key_5_cast_fp16 = add(x = pretrained_out_43_cast_fp16, y = lora_out_43_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")];
681 tensor<string, []> pretrained_out_45_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_45_pad_type_0"), val = tensor<string, []>("valid")];
682 tensor<int32, [2]> pretrained_out_45_strides_0 = const()[name = tensor<string, []>("pretrained_out_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
683 tensor<int32, [4]> pretrained_out_45_pad_0 = const()[name = tensor<string, []>("pretrained_out_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
684 tensor<int32, [2]> pretrained_out_45_dilations_0 = const()[name = tensor<string, []>("pretrained_out_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
685 tensor<int32, []> pretrained_out_45_groups_0 = const()[name = tensor<string, []>("pretrained_out_45_groups_0"), val = tensor<int32, []>(1)];
686 tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164179584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164998848))), name = tensor<string, []>("layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
687 tensor<fp16, [1280]> layers_2_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164998976)))];
688 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_45_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_45_dilations_0, groups = pretrained_out_45_groups_0, pad = pretrained_out_45_pad_0, pad_type = pretrained_out_45_pad_type_0, strides = pretrained_out_45_strides_0, weight = layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("pretrained_out_45_cast_fp16")];
689 tensor<string, []> input_65_pad_type_0 = const()[name = tensor<string, []>("input_65_pad_type_0"), val = tensor<string, []>("valid")];
690 tensor<int32, [2]> input_65_strides_0 = const()[name = tensor<string, []>("input_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
691 tensor<int32, [4]> input_65_pad_0 = const()[name = tensor<string, []>("input_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
692 tensor<int32, [2]> input_65_dilations_0 = const()[name = tensor<string, []>("input_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
693 tensor<int32, []> input_65_groups_0 = const()[name = tensor<string, []>("input_65_groups_0"), val = tensor<int32, []>(1)];
694 tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165001600)))];
695 tensor<fp16, [1, 16, 1, 1]> input_65_cast_fp16 = conv(dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = layers_2_self_attn_v_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
696 tensor<string, []> lora_out_45_pad_type_0 = const()[name = tensor<string, []>("lora_out_45_pad_type_0"), val = tensor<string, []>("valid")];
697 tensor<int32, [2]> lora_out_45_strides_0 = const()[name = tensor<string, []>("lora_out_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
698 tensor<int32, [4]> lora_out_45_pad_0 = const()[name = tensor<string, []>("lora_out_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
699 tensor<int32, [2]> lora_out_45_dilations_0 = const()[name = tensor<string, []>("lora_out_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
700 tensor<int32, []> lora_out_45_groups_0 = const()[name = tensor<string, []>("lora_out_45_groups_0"), val = tensor<int32, []>(1)];
701 tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165042624)))];
702 tensor<fp16, [1, 1280, 1, 1]> lora_out_45_cast_fp16 = conv(dilations = lora_out_45_dilations_0, groups = lora_out_45_groups_0, pad = lora_out_45_pad_0, pad_type = lora_out_45_pad_type_0, strides = lora_out_45_strides_0, weight = layers_2_self_attn_v_proj_loraB_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("lora_out_45_cast_fp16")];
703 tensor<fp16, [1, 1280, 1, 1]> current_value_5_cast_fp16 = add(x = pretrained_out_45_cast_fp16, y = lora_out_45_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")];
704 tensor<fp16, [1, 1280, 1, 448]> var_924_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_924_cast_fp16")];
705 tensor<fp16, [1, 1280, 1, 448]> var_926_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_177_cast_fp16)[name = tensor<string, []>("op_926_cast_fp16")];
706 tensor<fp16, [1, 1280, 1, 448]> key_9_cast_fp16 = add(x = var_924_cast_fp16, y = var_926_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
707 tensor<fp16, [1, 1280, 1, 448]> var_928_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_928_cast_fp16")];
708 tensor<fp16, [1, 1280, 1, 448]> var_930_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_177_cast_fp16)[name = tensor<string, []>("op_930_cast_fp16")];
709 tensor<fp16, [1, 1280, 1, 448]> value_9_cast_fp16 = add(x = var_928_cast_fp16, y = var_930_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
710 tensor<int32, [4]> var_933 = const()[name = tensor<string, []>("op_933"), val = tensor<int32, [4]>([1, 20, 64, -1])];
711 tensor<fp16, [1, 20, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_933, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
712 tensor<fp16, []> var_935_to_fp16 = const()[name = tensor<string, []>("op_935_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
713 tensor<fp16, [1, 20, 64, 1]> var_936_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_935_to_fp16)[name = tensor<string, []>("op_936_cast_fp16")];
714 tensor<int32, [4]> var_937 = const()[name = tensor<string, []>("op_937"), val = tensor<int32, [4]>([1, 20, 64, -1])];
715 tensor<fp16, [1, 20, 64, 448]> var_938_cast_fp16 = reshape(shape = var_937, x = key_9_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
716 tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
717 tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
718 tensor<fp16, [1, 20, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_936_cast_fp16, y = var_938_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
719 tensor<fp16, [1, 20, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
720 tensor<fp16, [1, 20, 1, 448]> var_946_cast_fp16 = softmax(axis = var_812, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_946_cast_fp16")];
721 tensor<int32, [4]> var_947 = const()[name = tensor<string, []>("op_947"), val = tensor<int32, [4]>([1, 20, 64, -1])];
722 tensor<fp16, [1, 20, 64, 448]> var_948_cast_fp16 = reshape(shape = var_947, x = value_9_cast_fp16)[name = tensor<string, []>("op_948_cast_fp16")];
723 tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
724 tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
725 tensor<fp16, [1, 20, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_948_cast_fp16, y = var_946_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
726 tensor<int32, [4]> var_951 = const()[name = tensor<string, []>("op_951"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
727 tensor<fp16, [1, 1280, 1, 1]> input_67_cast_fp16 = reshape(shape = var_951, x = attn_9_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
728 tensor<string, []> pretrained_out_47_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_47_pad_type_0"), val = tensor<string, []>("valid")];
729 tensor<int32, [2]> pretrained_out_47_strides_0 = const()[name = tensor<string, []>("pretrained_out_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
730 tensor<int32, [4]> pretrained_out_47_pad_0 = const()[name = tensor<string, []>("pretrained_out_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
731 tensor<int32, [2]> pretrained_out_47_dilations_0 = const()[name = tensor<string, []>("pretrained_out_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
732 tensor<int32, []> pretrained_out_47_groups_0 = const()[name = tensor<string, []>("pretrained_out_47_groups_0"), val = tensor<int32, []>(1)];
733 tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165083648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165902912))), name = tensor<string, []>("layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
734 tensor<fp16, [1280]> layers_2_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165903040)))];
735 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_47_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_47_dilations_0, groups = pretrained_out_47_groups_0, pad = pretrained_out_47_pad_0, pad_type = pretrained_out_47_pad_type_0, strides = pretrained_out_47_strides_0, weight = layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor<string, []>("pretrained_out_47_cast_fp16")];
736 tensor<string, []> input_69_pad_type_0 = const()[name = tensor<string, []>("input_69_pad_type_0"), val = tensor<string, []>("valid")];
737 tensor<int32, [2]> input_69_strides_0 = const()[name = tensor<string, []>("input_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
738 tensor<int32, [4]> input_69_pad_0 = const()[name = tensor<string, []>("input_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
739 tensor<int32, [2]> input_69_dilations_0 = const()[name = tensor<string, []>("input_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
740 tensor<int32, []> input_69_groups_0 = const()[name = tensor<string, []>("input_69_groups_0"), val = tensor<int32, []>(1)];
741 tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165905664)))];
742 tensor<fp16, [1, 16, 1, 1]> input_69_cast_fp16 = conv(dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_2_self_attn_o_proj_loraA_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
743 tensor<string, []> lora_out_47_pad_type_0 = const()[name = tensor<string, []>("lora_out_47_pad_type_0"), val = tensor<string, []>("valid")];
744 tensor<int32, [2]> lora_out_47_strides_0 = const()[name = tensor<string, []>("lora_out_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
745 tensor<int32, [4]> lora_out_47_pad_0 = const()[name = tensor<string, []>("lora_out_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
746 tensor<int32, [2]> lora_out_47_dilations_0 = const()[name = tensor<string, []>("lora_out_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
747 tensor<int32, []> lora_out_47_groups_0 = const()[name = tensor<string, []>("lora_out_47_groups_0"), val = tensor<int32, []>(1)];
748 tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165946688)))];
749 tensor<fp16, [1, 1280, 1, 1]> lora_out_47_cast_fp16 = conv(dilations = lora_out_47_dilations_0, groups = lora_out_47_groups_0, pad = lora_out_47_pad_0, pad_type = lora_out_47_pad_type_0, strides = lora_out_47_strides_0, weight = layers_2_self_attn_o_proj_loraB_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("lora_out_47_cast_fp16")];
750 tensor<fp16, [1, 1280, 1, 1]> obj_35_cast_fp16 = add(x = pretrained_out_47_cast_fp16, y = lora_out_47_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
751 tensor<fp16, [1, 1280, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
752 tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
753 tensor<fp16, []> var_989_to_fp16 = const()[name = tensor<string, []>("op_989_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
754 tensor<fp16, [1, 1280, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_989_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
755 tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165987712)))];
756 tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165990336)))];
757 tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
758 tensor<fp16, [1, 1280, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
759 tensor<string, []> pretrained_out_49_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_49_pad_type_0"), val = tensor<string, []>("valid")];
760 tensor<int32, [2]> pretrained_out_49_strides_0 = const()[name = tensor<string, []>("pretrained_out_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
761 tensor<int32, [4]> pretrained_out_49_pad_0 = const()[name = tensor<string, []>("pretrained_out_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
762 tensor<int32, [2]> pretrained_out_49_dilations_0 = const()[name = tensor<string, []>("pretrained_out_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
763 tensor<int32, []> pretrained_out_49_groups_0 = const()[name = tensor<string, []>("pretrained_out_49_groups_0"), val = tensor<int32, []>(1)];
764 tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165992960))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166812224))), name = tensor<string, []>("layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
765 tensor<fp16, [1280]> layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166812352)))];
766 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_49_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_49_dilations_0, groups = pretrained_out_49_groups_0, pad = pretrained_out_49_pad_0, pad_type = pretrained_out_49_pad_type_0, strides = pretrained_out_49_strides_0, weight = layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("pretrained_out_49_cast_fp16")];
767 tensor<string, []> input_71_pad_type_0 = const()[name = tensor<string, []>("input_71_pad_type_0"), val = tensor<string, []>("valid")];
768 tensor<int32, [2]> input_71_strides_0 = const()[name = tensor<string, []>("input_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
769 tensor<int32, [4]> input_71_pad_0 = const()[name = tensor<string, []>("input_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
770 tensor<int32, [2]> input_71_dilations_0 = const()[name = tensor<string, []>("input_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
771 tensor<int32, []> input_71_groups_0 = const()[name = tensor<string, []>("input_71_groups_0"), val = tensor<int32, []>(1)];
772 tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166814976)))];
773 tensor<fp16, [1, 16, 1, 1]> input_71_cast_fp16 = conv(dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = layers_2_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
774 tensor<string, []> lora_out_49_pad_type_0 = const()[name = tensor<string, []>("lora_out_49_pad_type_0"), val = tensor<string, []>("valid")];
775 tensor<int32, [2]> lora_out_49_strides_0 = const()[name = tensor<string, []>("lora_out_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
776 tensor<int32, [4]> lora_out_49_pad_0 = const()[name = tensor<string, []>("lora_out_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
777 tensor<int32, [2]> lora_out_49_dilations_0 = const()[name = tensor<string, []>("lora_out_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
778 tensor<int32, []> lora_out_49_groups_0 = const()[name = tensor<string, []>("lora_out_49_groups_0"), val = tensor<int32, []>(1)];
779 tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166856000)))];
780 tensor<fp16, [1, 1280, 1, 1]> lora_out_49_cast_fp16 = conv(dilations = lora_out_49_dilations_0, groups = lora_out_49_groups_0, pad = lora_out_49_pad_0, pad_type = lora_out_49_pad_type_0, strides = lora_out_49_strides_0, weight = layers_2_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("lora_out_49_cast_fp16")];
781 tensor<fp16, [1, 1280, 1, 1]> query_11_cast_fp16 = add(x = pretrained_out_49_cast_fp16, y = lora_out_49_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
782 tensor<string, []> pretrained_out_51_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_51_pad_type_0"), val = tensor<string, []>("valid")];
783 tensor<int32, [2]> pretrained_out_51_strides_0 = const()[name = tensor<string, []>("pretrained_out_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
784 tensor<int32, [4]> pretrained_out_51_pad_0 = const()[name = tensor<string, []>("pretrained_out_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
785 tensor<int32, [2]> pretrained_out_51_dilations_0 = const()[name = tensor<string, []>("pretrained_out_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
786 tensor<int32, []> pretrained_out_51_groups_0 = const()[name = tensor<string, []>("pretrained_out_51_groups_0"), val = tensor<int32, []>(1)];
787 tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166897024))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167716288))), name = tensor<string, []>("layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
788 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_51_cast_fp16 = conv(dilations = pretrained_out_51_dilations_0, groups = pretrained_out_51_groups_0, pad = pretrained_out_51_pad_0, pad_type = pretrained_out_51_pad_type_0, strides = pretrained_out_51_strides_0, weight = layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_51_cast_fp16")];
789 tensor<string, []> input_73_pad_type_0 = const()[name = tensor<string, []>("input_73_pad_type_0"), val = tensor<string, []>("valid")];
790 tensor<int32, [2]> input_73_strides_0 = const()[name = tensor<string, []>("input_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
791 tensor<int32, [4]> input_73_pad_0 = const()[name = tensor<string, []>("input_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
792 tensor<int32, [2]> input_73_dilations_0 = const()[name = tensor<string, []>("input_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
793 tensor<int32, []> input_73_groups_0 = const()[name = tensor<string, []>("input_73_groups_0"), val = tensor<int32, []>(1)];
794 tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167716416)))];
795 tensor<fp16, [1, 16, 1, 1500]> input_73_cast_fp16 = conv(dilations = input_73_dilations_0, groups = input_73_groups_0, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = input_73_strides_0, weight = layers_2_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_73_cast_fp16")];
796 tensor<string, []> lora_out_51_pad_type_0 = const()[name = tensor<string, []>("lora_out_51_pad_type_0"), val = tensor<string, []>("valid")];
797 tensor<int32, [2]> lora_out_51_strides_0 = const()[name = tensor<string, []>("lora_out_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
798 tensor<int32, [4]> lora_out_51_pad_0 = const()[name = tensor<string, []>("lora_out_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
799 tensor<int32, [2]> lora_out_51_dilations_0 = const()[name = tensor<string, []>("lora_out_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
800 tensor<int32, []> lora_out_51_groups_0 = const()[name = tensor<string, []>("lora_out_51_groups_0"), val = tensor<int32, []>(1)];
801 tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167757440)))];
802 tensor<fp16, [1, 1280, 1, 1500]> lora_out_51_cast_fp16 = conv(dilations = lora_out_51_dilations_0, groups = lora_out_51_groups_0, pad = lora_out_51_pad_0, pad_type = lora_out_51_pad_type_0, strides = lora_out_51_strides_0, weight = layers_2_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("lora_out_51_cast_fp16")];
803 tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = add(x = pretrained_out_51_cast_fp16, y = lora_out_51_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
804 tensor<string, []> pretrained_out_53_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_53_pad_type_0"), val = tensor<string, []>("valid")];
805 tensor<int32, [2]> pretrained_out_53_strides_0 = const()[name = tensor<string, []>("pretrained_out_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
806 tensor<int32, [4]> pretrained_out_53_pad_0 = const()[name = tensor<string, []>("pretrained_out_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
807 tensor<int32, [2]> pretrained_out_53_dilations_0 = const()[name = tensor<string, []>("pretrained_out_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
808 tensor<int32, []> pretrained_out_53_groups_0 = const()[name = tensor<string, []>("pretrained_out_53_groups_0"), val = tensor<int32, []>(1)];
809 tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167798464))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168617728))), name = tensor<string, []>("layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
810 tensor<fp16, [1280]> layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168617856)))];
811 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_53_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_53_dilations_0, groups = pretrained_out_53_groups_0, pad = pretrained_out_53_pad_0, pad_type = pretrained_out_53_pad_type_0, strides = pretrained_out_53_strides_0, weight = layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_53_cast_fp16")];
812 tensor<string, []> input_75_pad_type_0 = const()[name = tensor<string, []>("input_75_pad_type_0"), val = tensor<string, []>("valid")];
813 tensor<int32, [2]> input_75_strides_0 = const()[name = tensor<string, []>("input_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
814 tensor<int32, [4]> input_75_pad_0 = const()[name = tensor<string, []>("input_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
815 tensor<int32, [2]> input_75_dilations_0 = const()[name = tensor<string, []>("input_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
816 tensor<int32, []> input_75_groups_0 = const()[name = tensor<string, []>("input_75_groups_0"), val = tensor<int32, []>(1)];
817 tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168620480)))];
818 tensor<fp16, [1, 16, 1, 1500]> input_75_cast_fp16 = conv(dilations = input_75_dilations_0, groups = input_75_groups_0, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = input_75_strides_0, weight = layers_2_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_75_cast_fp16")];
819 tensor<string, []> lora_out_53_pad_type_0 = const()[name = tensor<string, []>("lora_out_53_pad_type_0"), val = tensor<string, []>("valid")];
820 tensor<int32, [2]> lora_out_53_strides_0 = const()[name = tensor<string, []>("lora_out_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
821 tensor<int32, [4]> lora_out_53_pad_0 = const()[name = tensor<string, []>("lora_out_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
822 tensor<int32, [2]> lora_out_53_dilations_0 = const()[name = tensor<string, []>("lora_out_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
823 tensor<int32, []> lora_out_53_groups_0 = const()[name = tensor<string, []>("lora_out_53_groups_0"), val = tensor<int32, []>(1)];
824 tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168661504)))];
825 tensor<fp16, [1, 1280, 1, 1500]> lora_out_53_cast_fp16 = conv(dilations = lora_out_53_dilations_0, groups = lora_out_53_groups_0, pad = lora_out_53_pad_0, pad_type = lora_out_53_pad_type_0, strides = lora_out_53_strides_0, weight = layers_2_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("lora_out_53_cast_fp16")];
826 tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = add(x = pretrained_out_53_cast_fp16, y = lora_out_53_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")];
827 tensor<int32, [4]> var_1072 = const()[name = tensor<string, []>("op_1072"), val = tensor<int32, [4]>([1, 20, 64, -1])];
828 tensor<fp16, [1, 20, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_1072, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
829 tensor<fp16, []> var_1074_to_fp16 = const()[name = tensor<string, []>("op_1074_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
830 tensor<fp16, [1, 20, 64, 1]> var_1075_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1074_to_fp16)[name = tensor<string, []>("op_1075_cast_fp16")];
831 tensor<int32, [4]> var_1076 = const()[name = tensor<string, []>("op_1076"), val = tensor<int32, [4]>([1, 20, 64, -1])];
832 tensor<fp16, [1, 20, 64, 1500]> var_1077_cast_fp16 = reshape(shape = var_1076, x = key_11_cast_fp16)[name = tensor<string, []>("op_1077_cast_fp16")];
833 tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
834 tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
835 tensor<fp16, [1, 20, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1075_cast_fp16, y = var_1077_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
836 tensor<fp16, [1, 20, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_812, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
837 tensor<int32, [4]> var_1081 = const()[name = tensor<string, []>("op_1081"), val = tensor<int32, [4]>([1, 20, 64, -1])];
838 tensor<fp16, [1, 20, 64, 1500]> var_1082_cast_fp16 = reshape(shape = var_1081, x = value_11_cast_fp16)[name = tensor<string, []>("op_1082_cast_fp16")];
839 tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
840 tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
841 tensor<fp16, [1, 20, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1082_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
842 tensor<int32, [4]> var_1085 = const()[name = tensor<string, []>("op_1085"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
843 tensor<fp16, [1, 1280, 1, 1]> input_77_cast_fp16 = reshape(shape = var_1085, x = attn_11_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
844 tensor<string, []> pretrained_out_55_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_55_pad_type_0"), val = tensor<string, []>("valid")];
845 tensor<int32, [2]> pretrained_out_55_strides_0 = const()[name = tensor<string, []>("pretrained_out_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
846 tensor<int32, [4]> pretrained_out_55_pad_0 = const()[name = tensor<string, []>("pretrained_out_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
847 tensor<int32, [2]> pretrained_out_55_dilations_0 = const()[name = tensor<string, []>("pretrained_out_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
848 tensor<int32, []> pretrained_out_55_groups_0 = const()[name = tensor<string, []>("pretrained_out_55_groups_0"), val = tensor<int32, []>(1)];
849 tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168702528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169521792))), name = tensor<string, []>("layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
850 tensor<fp16, [1280]> layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169521920)))];
851 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_55_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_55_dilations_0, groups = pretrained_out_55_groups_0, pad = pretrained_out_55_pad_0, pad_type = pretrained_out_55_pad_type_0, strides = pretrained_out_55_strides_0, weight = layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor<string, []>("pretrained_out_55_cast_fp16")];
852 tensor<string, []> input_79_pad_type_0 = const()[name = tensor<string, []>("input_79_pad_type_0"), val = tensor<string, []>("valid")];
853 tensor<int32, [2]> input_79_strides_0 = const()[name = tensor<string, []>("input_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
854 tensor<int32, [4]> input_79_pad_0 = const()[name = tensor<string, []>("input_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
855 tensor<int32, [2]> input_79_dilations_0 = const()[name = tensor<string, []>("input_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
856 tensor<int32, []> input_79_groups_0 = const()[name = tensor<string, []>("input_79_groups_0"), val = tensor<int32, []>(1)];
857 tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169524544)))];
858 tensor<fp16, [1, 16, 1, 1]> input_79_cast_fp16 = conv(dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = layers_2_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
859 tensor<string, []> lora_out_55_pad_type_0 = const()[name = tensor<string, []>("lora_out_55_pad_type_0"), val = tensor<string, []>("valid")];
860 tensor<int32, [2]> lora_out_55_strides_0 = const()[name = tensor<string, []>("lora_out_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
861 tensor<int32, [4]> lora_out_55_pad_0 = const()[name = tensor<string, []>("lora_out_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
862 tensor<int32, [2]> lora_out_55_dilations_0 = const()[name = tensor<string, []>("lora_out_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
863 tensor<int32, []> lora_out_55_groups_0 = const()[name = tensor<string, []>("lora_out_55_groups_0"), val = tensor<int32, []>(1)];
864 tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169565568)))];
865 tensor<fp16, [1, 1280, 1, 1]> lora_out_55_cast_fp16 = conv(dilations = lora_out_55_dilations_0, groups = lora_out_55_groups_0, pad = lora_out_55_pad_0, pad_type = lora_out_55_pad_type_0, strides = lora_out_55_strides_0, weight = layers_2_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("lora_out_55_cast_fp16")];
866 tensor<fp16, [1, 1280, 1, 1]> obj_39_cast_fp16 = add(x = pretrained_out_55_cast_fp16, y = lora_out_55_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
867 tensor<fp16, [1, 1280, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
868 tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
869 tensor<fp16, []> var_1122_to_fp16 = const()[name = tensor<string, []>("op_1122_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
870 tensor<fp16, [1, 1280, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1122_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
871 tensor<fp16, [1280]> input_81_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_81_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169606592)))];
872 tensor<fp16, [1280]> input_81_beta_0_to_fp16 = const()[name = tensor<string, []>("input_81_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169609216)))];
873 tensor<fp16, []> input_81_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_81_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
874 tensor<fp16, [1, 1280, 1, 1]> input_81_cast_fp16 = batch_norm(beta = input_81_beta_0_to_fp16, epsilon = input_81_epsilon_0_to_fp16, gamma = input_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
875 tensor<string, []> pretrained_out_57_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_57_pad_type_0"), val = tensor<string, []>("valid")];
876 tensor<int32, [2]> pretrained_out_57_strides_0 = const()[name = tensor<string, []>("pretrained_out_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
877 tensor<int32, [4]> pretrained_out_57_pad_0 = const()[name = tensor<string, []>("pretrained_out_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
878 tensor<int32, [2]> pretrained_out_57_dilations_0 = const()[name = tensor<string, []>("pretrained_out_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
879 tensor<int32, []> pretrained_out_57_groups_0 = const()[name = tensor<string, []>("pretrained_out_57_groups_0"), val = tensor<int32, []>(1)];
880 tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169611840))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172888704))), name = tensor<string, []>("layers_2_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
881 tensor<fp16, [5120]> layers_2_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172888832)))];
882 tensor<fp16, [1, 5120, 1, 1]> pretrained_out_57_cast_fp16 = conv(bias = layers_2_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_57_dilations_0, groups = pretrained_out_57_groups_0, pad = pretrained_out_57_pad_0, pad_type = pretrained_out_57_pad_type_0, strides = pretrained_out_57_strides_0, weight = layers_2_fc1_pretrained_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor<string, []>("pretrained_out_57_cast_fp16")];
883 tensor<string, []> input_83_pad_type_0 = const()[name = tensor<string, []>("input_83_pad_type_0"), val = tensor<string, []>("valid")];
884 tensor<int32, [2]> input_83_strides_0 = const()[name = tensor<string, []>("input_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
885 tensor<int32, [4]> input_83_pad_0 = const()[name = tensor<string, []>("input_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
886 tensor<int32, [2]> input_83_dilations_0 = const()[name = tensor<string, []>("input_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
887 tensor<int32, []> input_83_groups_0 = const()[name = tensor<string, []>("input_83_groups_0"), val = tensor<int32, []>(1)];
888 tensor<fp16, [16, 1280, 1, 1]> layers_2_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172899136)))];
889 tensor<fp16, [1, 16, 1, 1]> input_83_cast_fp16 = conv(dilations = input_83_dilations_0, groups = input_83_groups_0, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = input_83_strides_0, weight = layers_2_fc1_loraA_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
890 tensor<string, []> lora_out_57_pad_type_0 = const()[name = tensor<string, []>("lora_out_57_pad_type_0"), val = tensor<string, []>("valid")];
891 tensor<int32, [2]> lora_out_57_strides_0 = const()[name = tensor<string, []>("lora_out_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
892 tensor<int32, [4]> lora_out_57_pad_0 = const()[name = tensor<string, []>("lora_out_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
893 tensor<int32, [2]> lora_out_57_dilations_0 = const()[name = tensor<string, []>("lora_out_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
894 tensor<int32, []> lora_out_57_groups_0 = const()[name = tensor<string, []>("lora_out_57_groups_0"), val = tensor<int32, []>(1)];
895 tensor<fp16, [5120, 16, 1, 1]> layers_2_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172940160)))];
896 tensor<fp16, [1, 5120, 1, 1]> lora_out_57_cast_fp16 = conv(dilations = lora_out_57_dilations_0, groups = lora_out_57_groups_0, pad = lora_out_57_pad_0, pad_type = lora_out_57_pad_type_0, strides = lora_out_57_strides_0, weight = layers_2_fc1_loraB_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("lora_out_57_cast_fp16")];
897 tensor<fp16, [1, 5120, 1, 1]> input_85_cast_fp16 = add(x = pretrained_out_57_cast_fp16, y = lora_out_57_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
898 tensor<string, []> input_87_mode_0 = const()[name = tensor<string, []>("input_87_mode_0"), val = tensor<string, []>("EXACT")];
899 tensor<fp16, [1, 5120, 1, 1]> input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
900 tensor<string, []> pretrained_out_59_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_59_pad_type_0"), val = tensor<string, []>("valid")];
901 tensor<int32, [2]> pretrained_out_59_strides_0 = const()[name = tensor<string, []>("pretrained_out_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
902 tensor<int32, [4]> pretrained_out_59_pad_0 = const()[name = tensor<string, []>("pretrained_out_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
903 tensor<int32, [2]> pretrained_out_59_dilations_0 = const()[name = tensor<string, []>("pretrained_out_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
904 tensor<int32, []> pretrained_out_59_groups_0 = const()[name = tensor<string, []>("pretrained_out_59_groups_0"), val = tensor<int32, []>(1)];
905 tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173104064))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176380928))), name = tensor<string, []>("layers_2_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
906 tensor<fp16, [1280]> layers_2_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176381056)))];
907 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_59_cast_fp16 = conv(bias = layers_2_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_59_dilations_0, groups = pretrained_out_59_groups_0, pad = pretrained_out_59_pad_0, pad_type = pretrained_out_59_pad_type_0, strides = pretrained_out_59_strides_0, weight = layers_2_fc2_pretrained_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor<string, []>("pretrained_out_59_cast_fp16")];
908 tensor<string, []> input_89_pad_type_0 = const()[name = tensor<string, []>("input_89_pad_type_0"), val = tensor<string, []>("valid")];
909 tensor<int32, [2]> input_89_strides_0 = const()[name = tensor<string, []>("input_89_strides_0"), val = tensor<int32, [2]>([1, 1])];
910 tensor<int32, [4]> input_89_pad_0 = const()[name = tensor<string, []>("input_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
911 tensor<int32, [2]> input_89_dilations_0 = const()[name = tensor<string, []>("input_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
912 tensor<int32, []> input_89_groups_0 = const()[name = tensor<string, []>("input_89_groups_0"), val = tensor<int32, []>(1)];
913 tensor<fp16, [16, 5120, 1, 1]> layers_2_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176383680)))];
914 tensor<fp16, [1, 16, 1, 1]> input_89_cast_fp16 = conv(dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = layers_2_fc2_loraA_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
915 tensor<string, []> lora_out_59_pad_type_0 = const()[name = tensor<string, []>("lora_out_59_pad_type_0"), val = tensor<string, []>("valid")];
916 tensor<int32, [2]> lora_out_59_strides_0 = const()[name = tensor<string, []>("lora_out_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
917 tensor<int32, [4]> lora_out_59_pad_0 = const()[name = tensor<string, []>("lora_out_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
918 tensor<int32, [2]> lora_out_59_dilations_0 = const()[name = tensor<string, []>("lora_out_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
919 tensor<int32, []> lora_out_59_groups_0 = const()[name = tensor<string, []>("lora_out_59_groups_0"), val = tensor<int32, []>(1)];
920 tensor<fp16, [1280, 16, 1, 1]> layers_2_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176547584)))];
921 tensor<fp16, [1, 1280, 1, 1]> lora_out_59_cast_fp16 = conv(dilations = lora_out_59_dilations_0, groups = lora_out_59_groups_0, pad = lora_out_59_pad_0, pad_type = lora_out_59_pad_type_0, strides = lora_out_59_strides_0, weight = layers_2_fc2_loraB_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("lora_out_59_cast_fp16")];
922 tensor<fp16, [1, 1280, 1, 1]> hidden_states_7_cast_fp16 = add(x = pretrained_out_59_cast_fp16, y = lora_out_59_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
923 tensor<fp16, [1, 1280, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
924 tensor<int32, []> var_1190 = const()[name = tensor<string, []>("op_1190"), val = tensor<int32, []>(3)];
925 tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
926 tensor<fp16, []> var_1216_to_fp16 = const()[name = tensor<string, []>("op_1216_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
927 tensor<fp16, [1, 1280, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1216_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
928 tensor<fp16, [1280]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176588608)))];
929 tensor<fp16, [1280]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176591232)))];
930 tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
931 tensor<fp16, [1, 1280, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
932 tensor<string, []> pretrained_out_61_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_61_pad_type_0"), val = tensor<string, []>("valid")];
933 tensor<int32, [2]> pretrained_out_61_strides_0 = const()[name = tensor<string, []>("pretrained_out_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
934 tensor<int32, [4]> pretrained_out_61_pad_0 = const()[name = tensor<string, []>("pretrained_out_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
935 tensor<int32, [2]> pretrained_out_61_dilations_0 = const()[name = tensor<string, []>("pretrained_out_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
936 tensor<int32, []> pretrained_out_61_groups_0 = const()[name = tensor<string, []>("pretrained_out_61_groups_0"), val = tensor<int32, []>(1)];
937 tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176593856))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177413120))), name = tensor<string, []>("layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
938 tensor<fp16, [1280]> layers_3_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177413248)))];
939 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_61_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_61_dilations_0, groups = pretrained_out_61_groups_0, pad = pretrained_out_61_pad_0, pad_type = pretrained_out_61_pad_type_0, strides = pretrained_out_61_strides_0, weight = layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("pretrained_out_61_cast_fp16")];
940 tensor<string, []> input_91_pad_type_0 = const()[name = tensor<string, []>("input_91_pad_type_0"), val = tensor<string, []>("valid")];
941 tensor<int32, [2]> input_91_strides_0 = const()[name = tensor<string, []>("input_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
942 tensor<int32, [4]> input_91_pad_0 = const()[name = tensor<string, []>("input_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
943 tensor<int32, [2]> input_91_dilations_0 = const()[name = tensor<string, []>("input_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
944 tensor<int32, []> input_91_groups_0 = const()[name = tensor<string, []>("input_91_groups_0"), val = tensor<int32, []>(1)];
945 tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177415872)))];
946 tensor<fp16, [1, 16, 1, 1]> input_91_cast_fp16 = conv(dilations = input_91_dilations_0, groups = input_91_groups_0, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = input_91_strides_0, weight = layers_3_self_attn_q_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
947 tensor<string, []> lora_out_61_pad_type_0 = const()[name = tensor<string, []>("lora_out_61_pad_type_0"), val = tensor<string, []>("valid")];
948 tensor<int32, [2]> lora_out_61_strides_0 = const()[name = tensor<string, []>("lora_out_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
949 tensor<int32, [4]> lora_out_61_pad_0 = const()[name = tensor<string, []>("lora_out_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
950 tensor<int32, [2]> lora_out_61_dilations_0 = const()[name = tensor<string, []>("lora_out_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
951 tensor<int32, []> lora_out_61_groups_0 = const()[name = tensor<string, []>("lora_out_61_groups_0"), val = tensor<int32, []>(1)];
952 tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177456896)))];
953 tensor<fp16, [1, 1280, 1, 1]> lora_out_61_cast_fp16 = conv(dilations = lora_out_61_dilations_0, groups = lora_out_61_groups_0, pad = lora_out_61_pad_0, pad_type = lora_out_61_pad_type_0, strides = lora_out_61_strides_0, weight = layers_3_self_attn_q_proj_loraB_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("lora_out_61_cast_fp16")];
954 tensor<fp16, [1, 1280, 1, 1]> query_13_cast_fp16 = add(x = pretrained_out_61_cast_fp16, y = lora_out_61_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
955 tensor<string, []> pretrained_out_63_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_63_pad_type_0"), val = tensor<string, []>("valid")];
956 tensor<int32, [2]> pretrained_out_63_strides_0 = const()[name = tensor<string, []>("pretrained_out_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
957 tensor<int32, [4]> pretrained_out_63_pad_0 = const()[name = tensor<string, []>("pretrained_out_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
958 tensor<int32, [2]> pretrained_out_63_dilations_0 = const()[name = tensor<string, []>("pretrained_out_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
959 tensor<int32, []> pretrained_out_63_groups_0 = const()[name = tensor<string, []>("pretrained_out_63_groups_0"), val = tensor<int32, []>(1)];
960 tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177497920))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178317184))), name = tensor<string, []>("layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
961 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_63_cast_fp16 = conv(dilations = pretrained_out_63_dilations_0, groups = pretrained_out_63_groups_0, pad = pretrained_out_63_pad_0, pad_type = pretrained_out_63_pad_type_0, strides = pretrained_out_63_strides_0, weight = layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("pretrained_out_63_cast_fp16")];
962 tensor<string, []> input_93_pad_type_0 = const()[name = tensor<string, []>("input_93_pad_type_0"), val = tensor<string, []>("valid")];
963 tensor<int32, [2]> input_93_strides_0 = const()[name = tensor<string, []>("input_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
964 tensor<int32, [4]> input_93_pad_0 = const()[name = tensor<string, []>("input_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
965 tensor<int32, [2]> input_93_dilations_0 = const()[name = tensor<string, []>("input_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
966 tensor<int32, []> input_93_groups_0 = const()[name = tensor<string, []>("input_93_groups_0"), val = tensor<int32, []>(1)];
967 tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178317312)))];
968 tensor<fp16, [1, 16, 1, 1]> input_93_cast_fp16 = conv(dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_3_self_attn_k_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
969 tensor<string, []> lora_out_63_pad_type_0 = const()[name = tensor<string, []>("lora_out_63_pad_type_0"), val = tensor<string, []>("valid")];
970 tensor<int32, [2]> lora_out_63_strides_0 = const()[name = tensor<string, []>("lora_out_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
971 tensor<int32, [4]> lora_out_63_pad_0 = const()[name = tensor<string, []>("lora_out_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
972 tensor<int32, [2]> lora_out_63_dilations_0 = const()[name = tensor<string, []>("lora_out_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
973 tensor<int32, []> lora_out_63_groups_0 = const()[name = tensor<string, []>("lora_out_63_groups_0"), val = tensor<int32, []>(1)];
974 tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178358336)))];
975 tensor<fp16, [1, 1280, 1, 1]> lora_out_63_cast_fp16 = conv(dilations = lora_out_63_dilations_0, groups = lora_out_63_groups_0, pad = lora_out_63_pad_0, pad_type = lora_out_63_pad_type_0, strides = lora_out_63_strides_0, weight = layers_3_self_attn_k_proj_loraB_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("lora_out_63_cast_fp16")];
976 tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = add(x = pretrained_out_63_cast_fp16, y = lora_out_63_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
977 tensor<string, []> pretrained_out_65_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_65_pad_type_0"), val = tensor<string, []>("valid")];
978 tensor<int32, [2]> pretrained_out_65_strides_0 = const()[name = tensor<string, []>("pretrained_out_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
979 tensor<int32, [4]> pretrained_out_65_pad_0 = const()[name = tensor<string, []>("pretrained_out_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
980 tensor<int32, [2]> pretrained_out_65_dilations_0 = const()[name = tensor<string, []>("pretrained_out_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
981 tensor<int32, []> pretrained_out_65_groups_0 = const()[name = tensor<string, []>("pretrained_out_65_groups_0"), val = tensor<int32, []>(1)];
982 tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178399360))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179218624))), name = tensor<string, []>("layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
983 tensor<fp16, [1280]> layers_3_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179218752)))];
984 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_65_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_65_dilations_0, groups = pretrained_out_65_groups_0, pad = pretrained_out_65_pad_0, pad_type = pretrained_out_65_pad_type_0, strides = pretrained_out_65_strides_0, weight = layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("pretrained_out_65_cast_fp16")];
985 tensor<string, []> input_95_pad_type_0 = const()[name = tensor<string, []>("input_95_pad_type_0"), val = tensor<string, []>("valid")];
986 tensor<int32, [2]> input_95_strides_0 = const()[name = tensor<string, []>("input_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
987 tensor<int32, [4]> input_95_pad_0 = const()[name = tensor<string, []>("input_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
988 tensor<int32, [2]> input_95_dilations_0 = const()[name = tensor<string, []>("input_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
989 tensor<int32, []> input_95_groups_0 = const()[name = tensor<string, []>("input_95_groups_0"), val = tensor<int32, []>(1)];
990 tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179221376)))];
991 tensor<fp16, [1, 16, 1, 1]> input_95_cast_fp16 = conv(dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = layers_3_self_attn_v_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
992 tensor<string, []> lora_out_65_pad_type_0 = const()[name = tensor<string, []>("lora_out_65_pad_type_0"), val = tensor<string, []>("valid")];
993 tensor<int32, [2]> lora_out_65_strides_0 = const()[name = tensor<string, []>("lora_out_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
994 tensor<int32, [4]> lora_out_65_pad_0 = const()[name = tensor<string, []>("lora_out_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
995 tensor<int32, [2]> lora_out_65_dilations_0 = const()[name = tensor<string, []>("lora_out_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
996 tensor<int32, []> lora_out_65_groups_0 = const()[name = tensor<string, []>("lora_out_65_groups_0"), val = tensor<int32, []>(1)];
997 tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179262400)))];
998 tensor<fp16, [1, 1280, 1, 1]> lora_out_65_cast_fp16 = conv(dilations = lora_out_65_dilations_0, groups = lora_out_65_groups_0, pad = lora_out_65_pad_0, pad_type = lora_out_65_pad_type_0, strides = lora_out_65_strides_0, weight = layers_3_self_attn_v_proj_loraB_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("lora_out_65_cast_fp16")];
999 tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = add(x = pretrained_out_65_cast_fp16, y = lora_out_65_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
1000 tensor<fp16, [1, 1280, 1, 448]> var_1302_cast_fp16 = mul(x = current_key_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_1302_cast_fp16")];
1001 tensor<fp16, [1, 1280, 1, 448]> var_1304_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_177_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
1002 tensor<fp16, [1, 1280, 1, 448]> key_13_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1304_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
1003 tensor<fp16, [1, 1280, 1, 448]> var_1306_cast_fp16 = mul(x = current_value_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_1306_cast_fp16")];
1004 tensor<fp16, [1, 1280, 1, 448]> var_1308_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_177_cast_fp16)[name = tensor<string, []>("op_1308_cast_fp16")];
1005 tensor<fp16, [1, 1280, 1, 448]> value_13_cast_fp16 = add(x = var_1306_cast_fp16, y = var_1308_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
1006 tensor<int32, [4]> var_1311 = const()[name = tensor<string, []>("op_1311"), val = tensor<int32, [4]>([1, 20, 64, -1])];
1007 tensor<fp16, [1, 20, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_1311, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
1008 tensor<fp16, []> var_1313_to_fp16 = const()[name = tensor<string, []>("op_1313_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1009 tensor<fp16, [1, 20, 64, 1]> var_1314_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1313_to_fp16)[name = tensor<string, []>("op_1314_cast_fp16")];
1010 tensor<int32, [4]> var_1315 = const()[name = tensor<string, []>("op_1315"), val = tensor<int32, [4]>([1, 20, 64, -1])];
1011 tensor<fp16, [1, 20, 64, 448]> var_1316_cast_fp16 = reshape(shape = var_1315, x = key_13_cast_fp16)[name = tensor<string, []>("op_1316_cast_fp16")];
1012 tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
1013 tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
1014 tensor<fp16, [1, 20, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1314_cast_fp16, y = var_1316_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
1015 tensor<fp16, [1, 20, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
1016 tensor<fp16, [1, 20, 1, 448]> var_1324_cast_fp16 = softmax(axis = var_1190, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_1324_cast_fp16")];
1017 tensor<int32, [4]> var_1325 = const()[name = tensor<string, []>("op_1325"), val = tensor<int32, [4]>([1, 20, 64, -1])];
1018 tensor<fp16, [1, 20, 64, 448]> var_1326_cast_fp16 = reshape(shape = var_1325, x = value_13_cast_fp16)[name = tensor<string, []>("op_1326_cast_fp16")];
1019 tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
1020 tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
1021 tensor<fp16, [1, 20, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1326_cast_fp16, y = var_1324_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
1022 tensor<int32, [4]> var_1329 = const()[name = tensor<string, []>("op_1329"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
1023 tensor<fp16, [1, 1280, 1, 1]> input_97_cast_fp16 = reshape(shape = var_1329, x = attn_13_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
1024 tensor<string, []> pretrained_out_67_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_67_pad_type_0"), val = tensor<string, []>("valid")];
1025 tensor<int32, [2]> pretrained_out_67_strides_0 = const()[name = tensor<string, []>("pretrained_out_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
1026 tensor<int32, [4]> pretrained_out_67_pad_0 = const()[name = tensor<string, []>("pretrained_out_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1027 tensor<int32, [2]> pretrained_out_67_dilations_0 = const()[name = tensor<string, []>("pretrained_out_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1028 tensor<int32, []> pretrained_out_67_groups_0 = const()[name = tensor<string, []>("pretrained_out_67_groups_0"), val = tensor<int32, []>(1)];
1029 tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179303424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180122688))), name = tensor<string, []>("layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
1030 tensor<fp16, [1280]> layers_3_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180122816)))];
1031 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_67_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_67_dilations_0, groups = pretrained_out_67_groups_0, pad = pretrained_out_67_pad_0, pad_type = pretrained_out_67_pad_type_0, strides = pretrained_out_67_strides_0, weight = layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor<string, []>("pretrained_out_67_cast_fp16")];
1032 tensor<string, []> input_99_pad_type_0 = const()[name = tensor<string, []>("input_99_pad_type_0"), val = tensor<string, []>("valid")];
1033 tensor<int32, [2]> input_99_strides_0 = const()[name = tensor<string, []>("input_99_strides_0"), val = tensor<int32, [2]>([1, 1])];
1034 tensor<int32, [4]> input_99_pad_0 = const()[name = tensor<string, []>("input_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1035 tensor<int32, [2]> input_99_dilations_0 = const()[name = tensor<string, []>("input_99_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1036 tensor<int32, []> input_99_groups_0 = const()[name = tensor<string, []>("input_99_groups_0"), val = tensor<int32, []>(1)];
1037 tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180125440)))];
1038 tensor<fp16, [1, 16, 1, 1]> input_99_cast_fp16 = conv(dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = layers_3_self_attn_o_proj_loraA_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
1039 tensor<string, []> lora_out_67_pad_type_0 = const()[name = tensor<string, []>("lora_out_67_pad_type_0"), val = tensor<string, []>("valid")];
1040 tensor<int32, [2]> lora_out_67_strides_0 = const()[name = tensor<string, []>("lora_out_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
1041 tensor<int32, [4]> lora_out_67_pad_0 = const()[name = tensor<string, []>("lora_out_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1042 tensor<int32, [2]> lora_out_67_dilations_0 = const()[name = tensor<string, []>("lora_out_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1043 tensor<int32, []> lora_out_67_groups_0 = const()[name = tensor<string, []>("lora_out_67_groups_0"), val = tensor<int32, []>(1)];
1044 tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180166464)))];
1045 tensor<fp16, [1, 1280, 1, 1]> lora_out_67_cast_fp16 = conv(dilations = lora_out_67_dilations_0, groups = lora_out_67_groups_0, pad = lora_out_67_pad_0, pad_type = lora_out_67_pad_type_0, strides = lora_out_67_strides_0, weight = layers_3_self_attn_o_proj_loraB_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("lora_out_67_cast_fp16")];
1046 tensor<fp16, [1, 1280, 1, 1]> obj_49_cast_fp16 = add(x = pretrained_out_67_cast_fp16, y = lora_out_67_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
1047 tensor<fp16, [1, 1280, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
1048 tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
1049 tensor<fp16, []> var_1367_to_fp16 = const()[name = tensor<string, []>("op_1367_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1050 tensor<fp16, [1, 1280, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1367_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
1051 tensor<fp16, [1280]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180207488)))];
1052 tensor<fp16, [1280]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180210112)))];
1053 tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1054 tensor<fp16, [1, 1280, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
1055 tensor<string, []> pretrained_out_69_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_69_pad_type_0"), val = tensor<string, []>("valid")];
1056 tensor<int32, [2]> pretrained_out_69_strides_0 = const()[name = tensor<string, []>("pretrained_out_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
1057 tensor<int32, [4]> pretrained_out_69_pad_0 = const()[name = tensor<string, []>("pretrained_out_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1058 tensor<int32, [2]> pretrained_out_69_dilations_0 = const()[name = tensor<string, []>("pretrained_out_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1059 tensor<int32, []> pretrained_out_69_groups_0 = const()[name = tensor<string, []>("pretrained_out_69_groups_0"), val = tensor<int32, []>(1)];
1060 tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180212736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181032000))), name = tensor<string, []>("layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
1061 tensor<fp16, [1280]> layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181032128)))];
1062 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_69_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_69_dilations_0, groups = pretrained_out_69_groups_0, pad = pretrained_out_69_pad_0, pad_type = pretrained_out_69_pad_type_0, strides = pretrained_out_69_strides_0, weight = layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor<string, []>("pretrained_out_69_cast_fp16")];
1063 tensor<string, []> input_101_pad_type_0 = const()[name = tensor<string, []>("input_101_pad_type_0"), val = tensor<string, []>("valid")];
1064 tensor<int32, [2]> input_101_strides_0 = const()[name = tensor<string, []>("input_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
1065 tensor<int32, [4]> input_101_pad_0 = const()[name = tensor<string, []>("input_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1066 tensor<int32, [2]> input_101_dilations_0 = const()[name = tensor<string, []>("input_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1067 tensor<int32, []> input_101_groups_0 = const()[name = tensor<string, []>("input_101_groups_0"), val = tensor<int32, []>(1)];
1068 tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181034752)))];
1069 tensor<fp16, [1, 16, 1, 1]> input_101_cast_fp16 = conv(dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = layers_3_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
1070 tensor<string, []> lora_out_69_pad_type_0 = const()[name = tensor<string, []>("lora_out_69_pad_type_0"), val = tensor<string, []>("valid")];
1071 tensor<int32, [2]> lora_out_69_strides_0 = const()[name = tensor<string, []>("lora_out_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
1072 tensor<int32, [4]> lora_out_69_pad_0 = const()[name = tensor<string, []>("lora_out_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1073 tensor<int32, [2]> lora_out_69_dilations_0 = const()[name = tensor<string, []>("lora_out_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1074 tensor<int32, []> lora_out_69_groups_0 = const()[name = tensor<string, []>("lora_out_69_groups_0"), val = tensor<int32, []>(1)];
1075 tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181075776)))];
1076 tensor<fp16, [1, 1280, 1, 1]> lora_out_69_cast_fp16 = conv(dilations = lora_out_69_dilations_0, groups = lora_out_69_groups_0, pad = lora_out_69_pad_0, pad_type = lora_out_69_pad_type_0, strides = lora_out_69_strides_0, weight = layers_3_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("lora_out_69_cast_fp16")];
1077 tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = add(x = pretrained_out_69_cast_fp16, y = lora_out_69_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
1078 tensor<string, []> pretrained_out_71_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_71_pad_type_0"), val = tensor<string, []>("valid")];
1079 tensor<int32, [2]> pretrained_out_71_strides_0 = const()[name = tensor<string, []>("pretrained_out_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
1080 tensor<int32, [4]> pretrained_out_71_pad_0 = const()[name = tensor<string, []>("pretrained_out_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1081 tensor<int32, [2]> pretrained_out_71_dilations_0 = const()[name = tensor<string, []>("pretrained_out_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1082 tensor<int32, []> pretrained_out_71_groups_0 = const()[name = tensor<string, []>("pretrained_out_71_groups_0"), val = tensor<int32, []>(1)];
1083 tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181116800))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181936064))), name = tensor<string, []>("layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
1084 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_71_cast_fp16 = conv(dilations = pretrained_out_71_dilations_0, groups = pretrained_out_71_groups_0, pad = pretrained_out_71_pad_0, pad_type = pretrained_out_71_pad_type_0, strides = pretrained_out_71_strides_0, weight = layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_71_cast_fp16")];
1085 tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
1086 tensor<int32, [2]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
1087 tensor<int32, [4]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1088 tensor<int32, [2]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1089 tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
1090 tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181936192)))];
1091 tensor<fp16, [1, 16, 1, 1500]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = layers_3_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_103_cast_fp16")];
1092 tensor<string, []> lora_out_71_pad_type_0 = const()[name = tensor<string, []>("lora_out_71_pad_type_0"), val = tensor<string, []>("valid")];
1093 tensor<int32, [2]> lora_out_71_strides_0 = const()[name = tensor<string, []>("lora_out_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
1094 tensor<int32, [4]> lora_out_71_pad_0 = const()[name = tensor<string, []>("lora_out_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1095 tensor<int32, [2]> lora_out_71_dilations_0 = const()[name = tensor<string, []>("lora_out_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1096 tensor<int32, []> lora_out_71_groups_0 = const()[name = tensor<string, []>("lora_out_71_groups_0"), val = tensor<int32, []>(1)];
1097 tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181977216)))];
1098 tensor<fp16, [1, 1280, 1, 1500]> lora_out_71_cast_fp16 = conv(dilations = lora_out_71_dilations_0, groups = lora_out_71_groups_0, pad = lora_out_71_pad_0, pad_type = lora_out_71_pad_type_0, strides = lora_out_71_strides_0, weight = layers_3_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("lora_out_71_cast_fp16")];
1099 tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = add(x = pretrained_out_71_cast_fp16, y = lora_out_71_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
1100 tensor<string, []> pretrained_out_73_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_73_pad_type_0"), val = tensor<string, []>("valid")];
1101 tensor<int32, [2]> pretrained_out_73_strides_0 = const()[name = tensor<string, []>("pretrained_out_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
1102 tensor<int32, [4]> pretrained_out_73_pad_0 = const()[name = tensor<string, []>("pretrained_out_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1103 tensor<int32, [2]> pretrained_out_73_dilations_0 = const()[name = tensor<string, []>("pretrained_out_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1104 tensor<int32, []> pretrained_out_73_groups_0 = const()[name = tensor<string, []>("pretrained_out_73_groups_0"), val = tensor<int32, []>(1)];
1105 tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182018240))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182837504))), name = tensor<string, []>("layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
1106 tensor<fp16, [1280]> layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182837632)))];
1107 tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_73_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_73_dilations_0, groups = pretrained_out_73_groups_0, pad = pretrained_out_73_pad_0, pad_type = pretrained_out_73_pad_type_0, strides = pretrained_out_73_strides_0, weight = layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_73_cast_fp16")];
1108 tensor<string, []> input_105_pad_type_0 = const()[name = tensor<string, []>("input_105_pad_type_0"), val = tensor<string, []>("valid")];
1109 tensor<int32, [2]> input_105_strides_0 = const()[name = tensor<string, []>("input_105_strides_0"), val = tensor<int32, [2]>([1, 1])];
1110 tensor<int32, [4]> input_105_pad_0 = const()[name = tensor<string, []>("input_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1111 tensor<int32, [2]> input_105_dilations_0 = const()[name = tensor<string, []>("input_105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1112 tensor<int32, []> input_105_groups_0 = const()[name = tensor<string, []>("input_105_groups_0"), val = tensor<int32, []>(1)];
1113 tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182840256)))];
1114 tensor<fp16, [1, 16, 1, 1500]> input_105_cast_fp16 = conv(dilations = input_105_dilations_0, groups = input_105_groups_0, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = input_105_strides_0, weight = layers_3_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_105_cast_fp16")];
1115 tensor<string, []> lora_out_73_pad_type_0 = const()[name = tensor<string, []>("lora_out_73_pad_type_0"), val = tensor<string, []>("valid")];
1116 tensor<int32, [2]> lora_out_73_strides_0 = const()[name = tensor<string, []>("lora_out_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
1117 tensor<int32, [4]> lora_out_73_pad_0 = const()[name = tensor<string, []>("lora_out_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1118 tensor<int32, [2]> lora_out_73_dilations_0 = const()[name = tensor<string, []>("lora_out_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1119 tensor<int32, []> lora_out_73_groups_0 = const()[name = tensor<string, []>("lora_out_73_groups_0"), val = tensor<int32, []>(1)];
1120 tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182881280)))];
1121 tensor<fp16, [1, 1280, 1, 1500]> lora_out_73_cast_fp16 = conv(dilations = lora_out_73_dilations_0, groups = lora_out_73_groups_0, pad = lora_out_73_pad_0, pad_type = lora_out_73_pad_type_0, strides = lora_out_73_strides_0, weight = layers_3_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("lora_out_73_cast_fp16")];
1122 tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = add(x = pretrained_out_73_cast_fp16, y = lora_out_73_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
1123 tensor<int32, [4]> var_1450 = const()[name = tensor<string, []>("op_1450"), val = tensor<int32, [4]>([1, 20, 64, -1])];
1124 tensor<fp16, [1, 20, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_1450, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
1125 tensor<fp16, []> var_1452_to_fp16 = const()[name = tensor<string, []>("op_1452_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1126 tensor<fp16, [1, 20, 64, 1]> var_1453_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_1452_to_fp16)[name = tensor<string, []>("op_1453_cast_fp16")];
1127 tensor<int32, [4]> var_1454 = const()[name = tensor<string, []>("op_1454"), val = tensor<int32, [4]>([1, 20, 64, -1])];
1128 tensor<fp16, [1, 20, 64, 1500]> var_1455_cast_fp16 = reshape(shape = var_1454, x = key_cast_fp16)[name = tensor<string, []>("op_1455_cast_fp16")];
1129 tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
1130 tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
1131 tensor<fp16, [1, 20, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_1453_cast_fp16, y = var_1455_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
1132 tensor<fp16, [1, 20, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_1190, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
1133 tensor<int32, [4]> var_1459 = const()[name = tensor<string, []>("op_1459"), val = tensor<int32, [4]>([1, 20, 64, -1])];
1134 tensor<fp16, [1, 20, 64, 1500]> var_1460_cast_fp16 = reshape(shape = var_1459, x = value_cast_fp16)[name = tensor<string, []>("op_1460_cast_fp16")];
1135 tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
1136 tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
1137 tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_1460_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
1138 tensor<int32, [4]> var_1463 = const()[name = tensor<string, []>("op_1463"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
1139 tensor<fp16, [1, 1280, 1, 1]> input_107_cast_fp16 = reshape(shape = var_1463, x = attn_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
1140 tensor<string, []> pretrained_out_75_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_75_pad_type_0"), val = tensor<string, []>("valid")];
1141 tensor<int32, [2]> pretrained_out_75_strides_0 = const()[name = tensor<string, []>("pretrained_out_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
1142 tensor<int32, [4]> pretrained_out_75_pad_0 = const()[name = tensor<string, []>("pretrained_out_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1143 tensor<int32, [2]> pretrained_out_75_dilations_0 = const()[name = tensor<string, []>("pretrained_out_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1144 tensor<int32, []> pretrained_out_75_groups_0 = const()[name = tensor<string, []>("pretrained_out_75_groups_0"), val = tensor<int32, []>(1)];
1145 tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182922304))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183741568))), name = tensor<string, []>("layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
1146 tensor<fp16, [1280]> layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183741696)))];
1147 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_75_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_75_dilations_0, groups = pretrained_out_75_groups_0, pad = pretrained_out_75_pad_0, pad_type = pretrained_out_75_pad_type_0, strides = pretrained_out_75_strides_0, weight = layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor<string, []>("pretrained_out_75_cast_fp16")];
1148 tensor<string, []> input_109_pad_type_0 = const()[name = tensor<string, []>("input_109_pad_type_0"), val = tensor<string, []>("valid")];
1149 tensor<int32, [2]> input_109_strides_0 = const()[name = tensor<string, []>("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
1150 tensor<int32, [4]> input_109_pad_0 = const()[name = tensor<string, []>("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1151 tensor<int32, [2]> input_109_dilations_0 = const()[name = tensor<string, []>("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1152 tensor<int32, []> input_109_groups_0 = const()[name = tensor<string, []>("input_109_groups_0"), val = tensor<int32, []>(1)];
1153 tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183744320)))];
1154 tensor<fp16, [1, 16, 1, 1]> input_109_cast_fp16 = conv(dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = layers_3_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
1155 tensor<string, []> lora_out_75_pad_type_0 = const()[name = tensor<string, []>("lora_out_75_pad_type_0"), val = tensor<string, []>("valid")];
1156 tensor<int32, [2]> lora_out_75_strides_0 = const()[name = tensor<string, []>("lora_out_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
1157 tensor<int32, [4]> lora_out_75_pad_0 = const()[name = tensor<string, []>("lora_out_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1158 tensor<int32, [2]> lora_out_75_dilations_0 = const()[name = tensor<string, []>("lora_out_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1159 tensor<int32, []> lora_out_75_groups_0 = const()[name = tensor<string, []>("lora_out_75_groups_0"), val = tensor<int32, []>(1)];
1160 tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183785344)))];
1161 tensor<fp16, [1, 1280, 1, 1]> lora_out_75_cast_fp16 = conv(dilations = lora_out_75_dilations_0, groups = lora_out_75_groups_0, pad = lora_out_75_pad_0, pad_type = lora_out_75_pad_type_0, strides = lora_out_75_strides_0, weight = layers_3_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("lora_out_75_cast_fp16")];
1162 tensor<fp16, [1, 1280, 1, 1]> obj_53_cast_fp16 = add(x = pretrained_out_75_cast_fp16, y = lora_out_75_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
1163 tensor<fp16, [1, 1280, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
1164 tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
1165 tensor<fp16, []> var_1500_to_fp16 = const()[name = tensor<string, []>("op_1500_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1166 tensor<fp16, [1, 1280, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1500_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
1167 tensor<fp16, [1280]> input_111_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_111_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183826368)))];
1168 tensor<fp16, [1280]> input_111_beta_0_to_fp16 = const()[name = tensor<string, []>("input_111_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183828992)))];
1169 tensor<fp16, []> input_111_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_111_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1170 tensor<fp16, [1, 1280, 1, 1]> input_111_cast_fp16 = batch_norm(beta = input_111_beta_0_to_fp16, epsilon = input_111_epsilon_0_to_fp16, gamma = input_111_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
1171 tensor<string, []> pretrained_out_77_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_77_pad_type_0"), val = tensor<string, []>("valid")];
1172 tensor<int32, [2]> pretrained_out_77_strides_0 = const()[name = tensor<string, []>("pretrained_out_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
1173 tensor<int32, [4]> pretrained_out_77_pad_0 = const()[name = tensor<string, []>("pretrained_out_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1174 tensor<int32, [2]> pretrained_out_77_dilations_0 = const()[name = tensor<string, []>("pretrained_out_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1175 tensor<int32, []> pretrained_out_77_groups_0 = const()[name = tensor<string, []>("pretrained_out_77_groups_0"), val = tensor<int32, []>(1)];
1176 tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183831616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187108480))), name = tensor<string, []>("layers_3_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
1177 tensor<fp16, [5120]> layers_3_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187108608)))];
1178 tensor<fp16, [1, 5120, 1, 1]> pretrained_out_77_cast_fp16 = conv(bias = layers_3_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_77_dilations_0, groups = pretrained_out_77_groups_0, pad = pretrained_out_77_pad_0, pad_type = pretrained_out_77_pad_type_0, strides = pretrained_out_77_strides_0, weight = layers_3_fc1_pretrained_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor<string, []>("pretrained_out_77_cast_fp16")];
1179 tensor<string, []> input_113_pad_type_0 = const()[name = tensor<string, []>("input_113_pad_type_0"), val = tensor<string, []>("valid")];
1180 tensor<int32, [2]> input_113_strides_0 = const()[name = tensor<string, []>("input_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
1181 tensor<int32, [4]> input_113_pad_0 = const()[name = tensor<string, []>("input_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1182 tensor<int32, [2]> input_113_dilations_0 = const()[name = tensor<string, []>("input_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1183 tensor<int32, []> input_113_groups_0 = const()[name = tensor<string, []>("input_113_groups_0"), val = tensor<int32, []>(1)];
1184 tensor<fp16, [16, 1280, 1, 1]> layers_3_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187118912)))];
1185 tensor<fp16, [1, 16, 1, 1]> input_113_cast_fp16 = conv(dilations = input_113_dilations_0, groups = input_113_groups_0, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = input_113_strides_0, weight = layers_3_fc1_loraA_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
1186 tensor<string, []> lora_out_77_pad_type_0 = const()[name = tensor<string, []>("lora_out_77_pad_type_0"), val = tensor<string, []>("valid")];
1187 tensor<int32, [2]> lora_out_77_strides_0 = const()[name = tensor<string, []>("lora_out_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
1188 tensor<int32, [4]> lora_out_77_pad_0 = const()[name = tensor<string, []>("lora_out_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1189 tensor<int32, [2]> lora_out_77_dilations_0 = const()[name = tensor<string, []>("lora_out_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1190 tensor<int32, []> lora_out_77_groups_0 = const()[name = tensor<string, []>("lora_out_77_groups_0"), val = tensor<int32, []>(1)];
1191 tensor<fp16, [5120, 16, 1, 1]> layers_3_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187159936)))];
1192 tensor<fp16, [1, 5120, 1, 1]> lora_out_77_cast_fp16 = conv(dilations = lora_out_77_dilations_0, groups = lora_out_77_groups_0, pad = lora_out_77_pad_0, pad_type = lora_out_77_pad_type_0, strides = lora_out_77_strides_0, weight = layers_3_fc1_loraB_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("lora_out_77_cast_fp16")];
1193 tensor<fp16, [1, 5120, 1, 1]> input_115_cast_fp16 = add(x = pretrained_out_77_cast_fp16, y = lora_out_77_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
1194 tensor<string, []> input_117_mode_0 = const()[name = tensor<string, []>("input_117_mode_0"), val = tensor<string, []>("EXACT")];
1195 tensor<fp16, [1, 5120, 1, 1]> input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
1196 tensor<string, []> pretrained_out_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_pad_type_0"), val = tensor<string, []>("valid")];
1197 tensor<int32, [2]> pretrained_out_strides_0 = const()[name = tensor<string, []>("pretrained_out_strides_0"), val = tensor<int32, [2]>([1, 1])];
1198 tensor<int32, [4]> pretrained_out_pad_0 = const()[name = tensor<string, []>("pretrained_out_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1199 tensor<int32, [2]> pretrained_out_dilations_0 = const()[name = tensor<string, []>("pretrained_out_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1200 tensor<int32, []> pretrained_out_groups_0 = const()[name = tensor<string, []>("pretrained_out_groups_0"), val = tensor<int32, []>(1)];
1201 tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187323840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192239104))), name = tensor<string, []>("layers_3_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
1202 tensor<fp16, [1280]> layers_3_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192239296)))];
1203 tensor<fp16, [1, 1280, 1, 1]> pretrained_out_cast_fp16 = conv(bias = layers_3_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_dilations_0, groups = pretrained_out_groups_0, pad = pretrained_out_pad_0, pad_type = pretrained_out_pad_type_0, strides = pretrained_out_strides_0, weight = layers_3_fc2_pretrained_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor<string, []>("pretrained_out_cast_fp16")];
1204 tensor<string, []> input_pad_type_0 = const()[name = tensor<string, []>("input_pad_type_0"), val = tensor<string, []>("valid")];
1205 tensor<int32, [2]> input_strides_0 = const()[name = tensor<string, []>("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
1206 tensor<int32, [4]> input_pad_0 = const()[name = tensor<string, []>("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1207 tensor<int32, [2]> input_dilations_0 = const()[name = tensor<string, []>("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1208 tensor<int32, []> input_groups_0 = const()[name = tensor<string, []>("input_groups_0"), val = tensor<int32, []>(1)];
1209 tensor<fp16, [16, 5120, 1, 1]> layers_3_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192241920)))];
1210 tensor<fp16, [1, 16, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_3_fc2_loraA_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
1211 tensor<string, []> lora_out_pad_type_0 = const()[name = tensor<string, []>("lora_out_pad_type_0"), val = tensor<string, []>("valid")];
1212 tensor<int32, [2]> lora_out_strides_0 = const()[name = tensor<string, []>("lora_out_strides_0"), val = tensor<int32, [2]>([1, 1])];
1213 tensor<int32, [4]> lora_out_pad_0 = const()[name = tensor<string, []>("lora_out_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1214 tensor<int32, [2]> lora_out_dilations_0 = const()[name = tensor<string, []>("lora_out_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1215 tensor<int32, []> lora_out_groups_0 = const()[name = tensor<string, []>("lora_out_groups_0"), val = tensor<int32, []>(1)];
1216 tensor<fp16, [1280, 16, 1, 1]> layers_3_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192405824)))];
1217 tensor<fp16, [1, 1280, 1, 1]> lora_out_cast_fp16 = conv(dilations = lora_out_dilations_0, groups = lora_out_groups_0, pad = lora_out_pad_0, pad_type = lora_out_pad_type_0, strides = lora_out_strides_0, weight = layers_3_fc2_loraB_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("lora_out_cast_fp16")];
1218 tensor<fp16, [1, 1280, 1, 1]> hidden_states_9_cast_fp16 = add(x = pretrained_out_cast_fp16, y = lora_out_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
1219 tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
1220 tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
1221 tensor<fp16, []> var_1575_to_fp16 = const()[name = tensor<string, []>("op_1575_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1222 tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1575_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
1223 tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192446848)))];
1224 tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192449472)))];
1225 tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1226 tensor<fp16, [1, 1280, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
1227 tensor<int32, [1]> var_1586_axes_0 = const()[name = tensor<string, []>("op_1586_axes_0"), val = tensor<int32, [1]>([2])];
1228 tensor<fp16, [1, 1280, 1]> var_1586_cast_fp16 = squeeze(axes = var_1586_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_1586_cast_fp16")];
1229 tensor<int32, [3]> var_1589_perm_0 = const()[name = tensor<string, []>("op_1589_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
1230 tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192452096)))];
1231 tensor<fp16, [1, 1, 1280]> var_1589_cast_fp16 = transpose(perm = var_1589_perm_0, x = var_1586_cast_fp16)[name = tensor<string, []>("transpose_0")];
1232 tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_1589_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
1233 tensor<int32, []> var_1593 = const()[name = tensor<string, []>("op_1593"), val = tensor<int32, []>(1)];
1234 tensor<bool, []> obj_59_interleave_0 = const()[name = tensor<string, []>("obj_59_interleave_0"), val = tensor<bool, []>(false)];
1235 tensor<fp16, [1, 5120, 1, 1]> key_cache_updates = concat(axis = var_1593, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_59_cast_fp16")];
1236 tensor<int32, []> var_1596 = const()[name = tensor<string, []>("op_1596"), val = tensor<int32, []>(1)];
1237 tensor<bool, []> obj_61_interleave_0 = const()[name = tensor<string, []>("obj_61_interleave_0"), val = tensor<bool, []>(false)];
1238 tensor<fp16, [1, 5120, 1, 1]> value_cache_updates = concat(axis = var_1596, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_61_cast_fp16")];
1239 tensor<int32, [4]> var_1607_begin_0 = const()[name = tensor<string, []>("op_1607_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
1240 tensor<int32, [4]> var_1607_end_0 = const()[name = tensor<string, []>("op_1607_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])];
1241 tensor<bool, [4]> var_1607_end_mask_0 = const()[name = tensor<string, []>("op_1607_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1242 tensor<fp16, [1, 1, 1, 1500]> var_1607_cast_fp16 = slice_by_index(begin = var_1607_begin_0, end = var_1607_end_0, end_mask = var_1607_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1607_cast_fp16")];
1243 tensor<int32, [4]> var_1610_begin_0 = const()[name = tensor<string, []>("op_1610_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1244 tensor<int32, [4]> var_1610_end_0 = const()[name = tensor<string, []>("op_1610_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
1245 tensor<bool, [4]> var_1610_end_mask_0 = const()[name = tensor<string, []>("op_1610_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
1246 tensor<bool, [4]> var_1610_squeeze_mask_0 = const()[name = tensor<string, []>("op_1610_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
1247 tensor<fp16, [1, 1, 1500]> var_1610_cast_fp16 = slice_by_index(begin = var_1610_begin_0, end = var_1610_end_0, end_mask = var_1610_end_mask_0, squeeze_mask = var_1610_squeeze_mask_0, x = var_1607_cast_fp16)[name = tensor<string, []>("op_1610_cast_fp16")];
1248 tensor<int32, [4]> var_1625_begin_0 = const()[name = tensor<string, []>("op_1625_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])];
1249 tensor<int32, [4]> var_1625_end_0 = const()[name = tensor<string, []>("op_1625_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])];
1250 tensor<bool, [4]> var_1625_end_mask_0 = const()[name = tensor<string, []>("op_1625_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1251 tensor<fp16, [1, 1, 1, 1500]> var_1625_cast_fp16 = slice_by_index(begin = var_1625_begin_0, end = var_1625_end_0, end_mask = var_1625_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1625_cast_fp16")];
1252 tensor<int32, [4]> var_1628_begin_0 = const()[name = tensor<string, []>("op_1628_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1253 tensor<int32, [4]> var_1628_end_0 = const()[name = tensor<string, []>("op_1628_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
1254 tensor<bool, [4]> var_1628_end_mask_0 = const()[name = tensor<string, []>("op_1628_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
1255 tensor<bool, [4]> var_1628_squeeze_mask_0 = const()[name = tensor<string, []>("op_1628_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
1256 tensor<fp16, [1, 1, 1500]> var_1628_cast_fp16 = slice_by_index(begin = var_1628_begin_0, end = var_1628_end_0, end_mask = var_1628_end_mask_0, squeeze_mask = var_1628_squeeze_mask_0, x = var_1625_cast_fp16)[name = tensor<string, []>("op_1628_cast_fp16")];
1257 tensor<int32, [4]> var_1643_begin_0 = const()[name = tensor<string, []>("op_1643_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
1258 tensor<int32, [4]> var_1643_end_0 = const()[name = tensor<string, []>("op_1643_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])];
1259 tensor<bool, [4]> var_1643_end_mask_0 = const()[name = tensor<string, []>("op_1643_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1260 tensor<fp16, [1, 1, 1, 1500]> var_1643_cast_fp16 = slice_by_index(begin = var_1643_begin_0, end = var_1643_end_0, end_mask = var_1643_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1643_cast_fp16")];
1261 tensor<int32, [4]> var_1646_begin_0 = const()[name = tensor<string, []>("op_1646_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1262 tensor<int32, [4]> var_1646_end_0 = const()[name = tensor<string, []>("op_1646_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
1263 tensor<bool, [4]> var_1646_end_mask_0 = const()[name = tensor<string, []>("op_1646_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
1264 tensor<bool, [4]> var_1646_squeeze_mask_0 = const()[name = tensor<string, []>("op_1646_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
1265 tensor<fp16, [1, 1, 1500]> var_1646_cast_fp16 = slice_by_index(begin = var_1646_begin_0, end = var_1646_end_0, end_mask = var_1646_end_mask_0, squeeze_mask = var_1646_squeeze_mask_0, x = var_1643_cast_fp16)[name = tensor<string, []>("op_1646_cast_fp16")];
1266 tensor<int32, [4]> var_1661_begin_0 = const()[name = tensor<string, []>("op_1661_begin_0"), val = tensor<int32, [4]>([0, 6, 0, 0])];
1267 tensor<int32, [4]> var_1661_end_0 = const()[name = tensor<string, []>("op_1661_end_0"), val = tensor<int32, [4]>([1, 7, 1, 1500])];
1268 tensor<bool, [4]> var_1661_end_mask_0 = const()[name = tensor<string, []>("op_1661_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1269 tensor<fp16, [1, 1, 1, 1500]> var_1661_cast_fp16 = slice_by_index(begin = var_1661_begin_0, end = var_1661_end_0, end_mask = var_1661_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1661_cast_fp16")];
1270 tensor<int32, [4]> var_1664_begin_0 = const()[name = tensor<string, []>("op_1664_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1271 tensor<int32, [4]> var_1664_end_0 = const()[name = tensor<string, []>("op_1664_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
1272 tensor<bool, [4]> var_1664_end_mask_0 = const()[name = tensor<string, []>("op_1664_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
1273 tensor<bool, [4]> var_1664_squeeze_mask_0 = const()[name = tensor<string, []>("op_1664_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
1274 tensor<fp16, [1, 1, 1500]> var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, squeeze_mask = var_1664_squeeze_mask_0, x = var_1661_cast_fp16)[name = tensor<string, []>("op_1664_cast_fp16")];
1275 tensor<int32, [4]> var_1679_begin_0 = const()[name = tensor<string, []>("op_1679_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])];
1276 tensor<int32, [4]> var_1679_end_0 = const()[name = tensor<string, []>("op_1679_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])];
1277 tensor<bool, [4]> var_1679_end_mask_0 = const()[name = tensor<string, []>("op_1679_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1278 tensor<fp16, [1, 1, 1, 1500]> var_1679_cast_fp16 = slice_by_index(begin = var_1679_begin_0, end = var_1679_end_0, end_mask = var_1679_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1679_cast_fp16")];
1279 tensor<int32, [4]> var_1682_begin_0 = const()[name = tensor<string, []>("op_1682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1280 tensor<int32, [4]> var_1682_end_0 = const()[name = tensor<string, []>("op_1682_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
1281 tensor<bool, [4]> var_1682_end_mask_0 = const()[name = tensor<string, []>("op_1682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
1282 tensor<bool, [4]> var_1682_squeeze_mask_0 = const()[name = tensor<string, []>("op_1682_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
1283 tensor<fp16, [1, 1, 1500]> var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, squeeze_mask = var_1682_squeeze_mask_0, x = var_1679_cast_fp16)[name = tensor<string, []>("op_1682_cast_fp16")];
1284 tensor<int32, [4]> var_1697_begin_0 = const()[name = tensor<string, []>("op_1697_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])];
1285 tensor<int32, [4]> var_1697_end_0 = const()[name = tensor<string, []>("op_1697_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])];
1286 tensor<bool, [4]> var_1697_end_mask_0 = const()[name = tensor<string, []>("op_1697_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1287 tensor<fp16, [1, 1, 1, 1500]> var_1697_cast_fp16 = slice_by_index(begin = var_1697_begin_0, end = var_1697_end_0, end_mask = var_1697_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1697_cast_fp16")];
1288 tensor<int32, [4]> var_1700_begin_0 = const()[name = tensor<string, []>("op_1700_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1289 tensor<int32, [4]> var_1700_end_0 = const()[name = tensor<string, []>("op_1700_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
1290 tensor<bool, [4]> var_1700_end_mask_0 = const()[name = tensor<string, []>("op_1700_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
1291 tensor<bool, [4]> var_1700_squeeze_mask_0 = const()[name = tensor<string, []>("op_1700_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
1292 tensor<fp16, [1, 1, 1500]> var_1700_cast_fp16 = slice_by_index(begin = var_1700_begin_0, end = var_1700_end_0, end_mask = var_1700_end_mask_0, squeeze_mask = var_1700_squeeze_mask_0, x = var_1697_cast_fp16)[name = tensor<string, []>("op_1700_cast_fp16")];
1293 tensor<int32, []> var_1707 = const()[name = tensor<string, []>("op_1707"), val = tensor<int32, []>(1)];
1294 tensor<bool, []> var_1708_interleave_0 = const()[name = tensor<string, []>("op_1708_interleave_0"), val = tensor<bool, []>(false)];
1295 tensor<fp16, [1, 6, 1500]> var_1708_cast_fp16 = concat(axis = var_1707, interleave = var_1708_interleave_0, values = (var_1610_cast_fp16, var_1628_cast_fp16, var_1646_cast_fp16, var_1664_cast_fp16, var_1682_cast_fp16, var_1700_cast_fp16))[name = tensor<string, []>("op_1708_cast_fp16")];
1296 tensor<bool, []> var_1711 = const()[name = tensor<string, []>("op_1711"), val = tensor<bool, []>(false)];
1297 tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])];
1298 tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_1711, x = var_1708_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
1299 } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
1300 }