openai_whisper-tiny.en/AudioEncoder.mlmodelc/model.mil
301.5 KB · 1713 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"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
3 {
4 func main<ios16>(tensor<fp16, [1, 80, 1, 3000]> melspectrogram_features) {
5 tensor<string, []> var_50_pad_type_0 = const()[name = tensor<string, []>("op_50_pad_type_0"), val = tensor<string, []>("custom")];
6 tensor<int32, [4]> var_50_pad_0 = const()[name = tensor<string, []>("op_50_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
7 tensor<int32, [2]> var_50_strides_0 = const()[name = tensor<string, []>("op_50_strides_0"), val = tensor<int32, [2]>([1, 1])];
8 tensor<int32, [2]> var_50_dilations_0 = const()[name = tensor<string, []>("op_50_dilations_0"), val = tensor<int32, [2]>([1, 1])];
9 tensor<int32, []> var_50_groups_0 = const()[name = tensor<string, []>("op_50_groups_0"), val = tensor<int32, []>(1)];
10 tensor<fp16, [384, 80, 1, 3]> var_25_to_fp16 = const()[name = tensor<string, []>("op_25_to_fp16"), val = tensor<fp16, [384, 80, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
11 tensor<fp16, [384]> var_31_to_fp16 = const()[name = tensor<string, []>("op_31_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184448)))];
12 tensor<fp16, [1, 384, 1, 3000]> var_50_cast_fp16 = conv(bias = var_31_to_fp16, dilations = var_50_dilations_0, groups = var_50_groups_0, pad = var_50_pad_0, pad_type = var_50_pad_type_0, strides = var_50_strides_0, weight = var_25_to_fp16, x = melspectrogram_features)[name = tensor<string, []>("op_50_cast_fp16")];
13 tensor<string, []> hidden_states_1_mode_0 = const()[name = tensor<string, []>("hidden_states_1_mode_0"), val = tensor<string, []>("EXACT")];
14 tensor<fp16, [1, 384, 1, 3000]> hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_50_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
15 tensor<string, []> var_90_pad_type_0 = const()[name = tensor<string, []>("op_90_pad_type_0"), val = tensor<string, []>("custom")];
16 tensor<int32, [4]> var_90_pad_0 = const()[name = tensor<string, []>("op_90_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
17 tensor<int32, [2]> var_90_strides_0 = const()[name = tensor<string, []>("op_90_strides_0"), val = tensor<int32, [2]>([2, 2])];
18 tensor<int32, [2]> var_90_dilations_0 = const()[name = tensor<string, []>("op_90_dilations_0"), val = tensor<int32, [2]>([1, 1])];
19 tensor<int32, []> var_90_groups_0 = const()[name = tensor<string, []>("op_90_groups_0"), val = tensor<int32, []>(1)];
20 tensor<fp16, [384, 384, 1, 3]> var_65_to_fp16 = const()[name = tensor<string, []>("op_65_to_fp16"), val = tensor<fp16, [384, 384, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185280)))];
21 tensor<fp16, [384]> var_71_to_fp16 = const()[name = tensor<string, []>("op_71_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1070080)))];
22 tensor<fp16, [1, 384, 1, 1500]> var_90_cast_fp16 = conv(bias = var_71_to_fp16, dilations = var_90_dilations_0, groups = var_90_groups_0, pad = var_90_pad_0, pad_type = var_90_pad_type_0, strides = var_90_strides_0, weight = var_65_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_90_cast_fp16")];
23 tensor<string, []> hidden_states_3_mode_0 = const()[name = tensor<string, []>("hidden_states_3_mode_0"), val = tensor<string, []>("EXACT")];
24 tensor<fp16, [1, 384, 1, 1500]> hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_90_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
25 tensor<fp16, [1, 384, 1, 1500]> var_108_to_fp16 = const()[name = tensor<string, []>("op_108_to_fp16"), val = tensor<fp16, [1, 384, 1, 1500]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1070912)))];
26 tensor<fp16, [1, 384, 1, 1500]> inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_108_to_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
27 tensor<int32, []> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, []>(3)];
28 tensor<int32, []> var_129 = const()[name = tensor<string, []>("op_129"), val = tensor<int32, []>(1)];
29 tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
30 tensor<fp16, []> var_146_to_fp16 = const()[name = tensor<string, []>("op_146_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
31 tensor<fp16, [1, 384, 1, 1500]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_146_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
32 tensor<fp16, [384]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2222976)))];
33 tensor<fp16, [384]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2223808)))];
34 tensor<fp16, [384]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2224640)))];
35 tensor<fp16, [384]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2225472)))];
36 tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
37 tensor<fp16, [1, 384, 1, 1500]> 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")];
38 tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")];
39 tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
40 tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
41 tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
42 tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)];
43 tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2226304)))];
44 tensor<fp16, [384]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2521280)))];
45 tensor<fp16, [1, 384, 1, 1500]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
46 tensor<string, []> key_1_pad_type_0 = const()[name = tensor<string, []>("key_1_pad_type_0"), val = tensor<string, []>("valid")];
47 tensor<int32, [2]> key_1_strides_0 = const()[name = tensor<string, []>("key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
48 tensor<int32, [4]> key_1_pad_0 = const()[name = tensor<string, []>("key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
49 tensor<int32, [2]> key_1_dilations_0 = const()[name = tensor<string, []>("key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
50 tensor<int32, []> key_1_groups_0 = const()[name = tensor<string, []>("key_1_groups_0"), val = tensor<int32, []>(1)];
51 tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2522112)))];
52 tensor<fp16, [1, 384, 1, 1500]> key_1_cast_fp16 = conv(dilations = key_1_dilations_0, groups = key_1_groups_0, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
53 tensor<string, []> value_1_pad_type_0 = const()[name = tensor<string, []>("value_1_pad_type_0"), val = tensor<string, []>("valid")];
54 tensor<int32, [2]> value_1_strides_0 = const()[name = tensor<string, []>("value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
55 tensor<int32, [4]> value_1_pad_0 = const()[name = tensor<string, []>("value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
56 tensor<int32, [2]> value_1_dilations_0 = const()[name = tensor<string, []>("value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
57 tensor<int32, []> value_1_groups_0 = const()[name = tensor<string, []>("value_1_groups_0"), val = tensor<int32, []>(1)];
58 tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2817088)))];
59 tensor<fp16, [384]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3112064)))];
60 tensor<fp16, [1, 384, 1, 1500]> value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
61 tensor<int32, [4]> var_184_begin_0 = const()[name = tensor<string, []>("op_184_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
62 tensor<int32, [4]> var_184_end_0 = const()[name = tensor<string, []>("op_184_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
63 tensor<bool, [4]> var_184_end_mask_0 = const()[name = tensor<string, []>("op_184_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
64 tensor<fp16, [1, 64, 1, 1500]> var_184_cast_fp16 = slice_by_index(begin = var_184_begin_0, end = var_184_end_0, end_mask = var_184_end_mask_0, x = query_1_cast_fp16)[name = tensor<string, []>("op_184_cast_fp16")];
65 tensor<int32, [4]> var_188_begin_0 = const()[name = tensor<string, []>("op_188_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
66 tensor<int32, [4]> var_188_end_0 = const()[name = tensor<string, []>("op_188_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
67 tensor<bool, [4]> var_188_end_mask_0 = const()[name = tensor<string, []>("op_188_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
68 tensor<fp16, [1, 64, 1, 1500]> var_188_cast_fp16 = slice_by_index(begin = var_188_begin_0, end = var_188_end_0, end_mask = var_188_end_mask_0, x = query_1_cast_fp16)[name = tensor<string, []>("op_188_cast_fp16")];
69 tensor<int32, [4]> var_192_begin_0 = const()[name = tensor<string, []>("op_192_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
70 tensor<int32, [4]> var_192_end_0 = const()[name = tensor<string, []>("op_192_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
71 tensor<bool, [4]> var_192_end_mask_0 = const()[name = tensor<string, []>("op_192_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
72 tensor<fp16, [1, 64, 1, 1500]> var_192_cast_fp16 = slice_by_index(begin = var_192_begin_0, end = var_192_end_0, end_mask = var_192_end_mask_0, x = query_1_cast_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
73 tensor<int32, [4]> var_196_begin_0 = const()[name = tensor<string, []>("op_196_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
74 tensor<int32, [4]> var_196_end_0 = const()[name = tensor<string, []>("op_196_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
75 tensor<bool, [4]> var_196_end_mask_0 = const()[name = tensor<string, []>("op_196_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
76 tensor<fp16, [1, 64, 1, 1500]> var_196_cast_fp16 = slice_by_index(begin = var_196_begin_0, end = var_196_end_0, end_mask = var_196_end_mask_0, x = query_1_cast_fp16)[name = tensor<string, []>("op_196_cast_fp16")];
77 tensor<int32, [4]> var_200_begin_0 = const()[name = tensor<string, []>("op_200_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
78 tensor<int32, [4]> var_200_end_0 = const()[name = tensor<string, []>("op_200_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
79 tensor<bool, [4]> var_200_end_mask_0 = const()[name = tensor<string, []>("op_200_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
80 tensor<fp16, [1, 64, 1, 1500]> var_200_cast_fp16 = slice_by_index(begin = var_200_begin_0, end = var_200_end_0, end_mask = var_200_end_mask_0, x = query_1_cast_fp16)[name = tensor<string, []>("op_200_cast_fp16")];
81 tensor<int32, [4]> var_204_begin_0 = const()[name = tensor<string, []>("op_204_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
82 tensor<int32, [4]> var_204_end_0 = const()[name = tensor<string, []>("op_204_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
83 tensor<bool, [4]> var_204_end_mask_0 = const()[name = tensor<string, []>("op_204_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
84 tensor<fp16, [1, 64, 1, 1500]> var_204_cast_fp16 = slice_by_index(begin = var_204_begin_0, end = var_204_end_0, end_mask = var_204_end_mask_0, x = query_1_cast_fp16)[name = tensor<string, []>("op_204_cast_fp16")];
85 tensor<int32, [4]> var_213_begin_0 = const()[name = tensor<string, []>("op_213_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
86 tensor<int32, [4]> var_213_end_0 = const()[name = tensor<string, []>("op_213_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
87 tensor<bool, [4]> var_213_end_mask_0 = const()[name = tensor<string, []>("op_213_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
88 tensor<fp16, [1, 64, 1, 375]> var_213_cast_fp16 = slice_by_index(begin = var_213_begin_0, end = var_213_end_0, end_mask = var_213_end_mask_0, x = var_184_cast_fp16)[name = tensor<string, []>("op_213_cast_fp16")];
89 tensor<int32, [4]> var_220_begin_0 = const()[name = tensor<string, []>("op_220_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
90 tensor<int32, [4]> var_220_end_0 = const()[name = tensor<string, []>("op_220_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
91 tensor<bool, [4]> var_220_end_mask_0 = const()[name = tensor<string, []>("op_220_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
92 tensor<fp16, [1, 64, 1, 375]> var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, x = var_184_cast_fp16)[name = tensor<string, []>("op_220_cast_fp16")];
93 tensor<int32, [4]> var_227_begin_0 = const()[name = tensor<string, []>("op_227_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
94 tensor<int32, [4]> var_227_end_0 = const()[name = tensor<string, []>("op_227_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
95 tensor<bool, [4]> var_227_end_mask_0 = const()[name = tensor<string, []>("op_227_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
96 tensor<fp16, [1, 64, 1, 375]> var_227_cast_fp16 = slice_by_index(begin = var_227_begin_0, end = var_227_end_0, end_mask = var_227_end_mask_0, x = var_184_cast_fp16)[name = tensor<string, []>("op_227_cast_fp16")];
97 tensor<int32, [4]> var_234_begin_0 = const()[name = tensor<string, []>("op_234_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
98 tensor<int32, [4]> var_234_end_0 = const()[name = tensor<string, []>("op_234_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
99 tensor<bool, [4]> var_234_end_mask_0 = const()[name = tensor<string, []>("op_234_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
100 tensor<fp16, [1, 64, 1, 375]> var_234_cast_fp16 = slice_by_index(begin = var_234_begin_0, end = var_234_end_0, end_mask = var_234_end_mask_0, x = var_184_cast_fp16)[name = tensor<string, []>("op_234_cast_fp16")];
101 tensor<int32, [4]> var_241_begin_0 = const()[name = tensor<string, []>("op_241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
102 tensor<int32, [4]> var_241_end_0 = const()[name = tensor<string, []>("op_241_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
103 tensor<bool, [4]> var_241_end_mask_0 = const()[name = tensor<string, []>("op_241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
104 tensor<fp16, [1, 64, 1, 375]> var_241_cast_fp16 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = var_188_cast_fp16)[name = tensor<string, []>("op_241_cast_fp16")];
105 tensor<int32, [4]> var_248_begin_0 = const()[name = tensor<string, []>("op_248_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
106 tensor<int32, [4]> var_248_end_0 = const()[name = tensor<string, []>("op_248_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
107 tensor<bool, [4]> var_248_end_mask_0 = const()[name = tensor<string, []>("op_248_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
108 tensor<fp16, [1, 64, 1, 375]> var_248_cast_fp16 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = var_188_cast_fp16)[name = tensor<string, []>("op_248_cast_fp16")];
109 tensor<int32, [4]> var_255_begin_0 = const()[name = tensor<string, []>("op_255_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
110 tensor<int32, [4]> var_255_end_0 = const()[name = tensor<string, []>("op_255_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
111 tensor<bool, [4]> var_255_end_mask_0 = const()[name = tensor<string, []>("op_255_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
112 tensor<fp16, [1, 64, 1, 375]> var_255_cast_fp16 = slice_by_index(begin = var_255_begin_0, end = var_255_end_0, end_mask = var_255_end_mask_0, x = var_188_cast_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
113 tensor<int32, [4]> var_262_begin_0 = const()[name = tensor<string, []>("op_262_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
114 tensor<int32, [4]> var_262_end_0 = const()[name = tensor<string, []>("op_262_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
115 tensor<bool, [4]> var_262_end_mask_0 = const()[name = tensor<string, []>("op_262_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
116 tensor<fp16, [1, 64, 1, 375]> var_262_cast_fp16 = slice_by_index(begin = var_262_begin_0, end = var_262_end_0, end_mask = var_262_end_mask_0, x = var_188_cast_fp16)[name = tensor<string, []>("op_262_cast_fp16")];
117 tensor<int32, [4]> var_269_begin_0 = const()[name = tensor<string, []>("op_269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
118 tensor<int32, [4]> var_269_end_0 = const()[name = tensor<string, []>("op_269_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
119 tensor<bool, [4]> var_269_end_mask_0 = const()[name = tensor<string, []>("op_269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
120 tensor<fp16, [1, 64, 1, 375]> var_269_cast_fp16 = slice_by_index(begin = var_269_begin_0, end = var_269_end_0, end_mask = var_269_end_mask_0, x = var_192_cast_fp16)[name = tensor<string, []>("op_269_cast_fp16")];
121 tensor<int32, [4]> var_276_begin_0 = const()[name = tensor<string, []>("op_276_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
122 tensor<int32, [4]> var_276_end_0 = const()[name = tensor<string, []>("op_276_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
123 tensor<bool, [4]> var_276_end_mask_0 = const()[name = tensor<string, []>("op_276_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
124 tensor<fp16, [1, 64, 1, 375]> var_276_cast_fp16 = slice_by_index(begin = var_276_begin_0, end = var_276_end_0, end_mask = var_276_end_mask_0, x = var_192_cast_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
125 tensor<int32, [4]> var_283_begin_0 = const()[name = tensor<string, []>("op_283_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
126 tensor<int32, [4]> var_283_end_0 = const()[name = tensor<string, []>("op_283_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
127 tensor<bool, [4]> var_283_end_mask_0 = const()[name = tensor<string, []>("op_283_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
128 tensor<fp16, [1, 64, 1, 375]> var_283_cast_fp16 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = var_192_cast_fp16)[name = tensor<string, []>("op_283_cast_fp16")];
129 tensor<int32, [4]> var_290_begin_0 = const()[name = tensor<string, []>("op_290_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
130 tensor<int32, [4]> var_290_end_0 = const()[name = tensor<string, []>("op_290_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
131 tensor<bool, [4]> var_290_end_mask_0 = const()[name = tensor<string, []>("op_290_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
132 tensor<fp16, [1, 64, 1, 375]> var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = var_192_cast_fp16)[name = tensor<string, []>("op_290_cast_fp16")];
133 tensor<int32, [4]> var_297_begin_0 = const()[name = tensor<string, []>("op_297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
134 tensor<int32, [4]> var_297_end_0 = const()[name = tensor<string, []>("op_297_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
135 tensor<bool, [4]> var_297_end_mask_0 = const()[name = tensor<string, []>("op_297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
136 tensor<fp16, [1, 64, 1, 375]> var_297_cast_fp16 = slice_by_index(begin = var_297_begin_0, end = var_297_end_0, end_mask = var_297_end_mask_0, x = var_196_cast_fp16)[name = tensor<string, []>("op_297_cast_fp16")];
137 tensor<int32, [4]> var_304_begin_0 = const()[name = tensor<string, []>("op_304_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
138 tensor<int32, [4]> var_304_end_0 = const()[name = tensor<string, []>("op_304_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
139 tensor<bool, [4]> var_304_end_mask_0 = const()[name = tensor<string, []>("op_304_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
140 tensor<fp16, [1, 64, 1, 375]> var_304_cast_fp16 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = var_196_cast_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
141 tensor<int32, [4]> var_311_begin_0 = const()[name = tensor<string, []>("op_311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
142 tensor<int32, [4]> var_311_end_0 = const()[name = tensor<string, []>("op_311_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
143 tensor<bool, [4]> var_311_end_mask_0 = const()[name = tensor<string, []>("op_311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
144 tensor<fp16, [1, 64, 1, 375]> var_311_cast_fp16 = slice_by_index(begin = var_311_begin_0, end = var_311_end_0, end_mask = var_311_end_mask_0, x = var_196_cast_fp16)[name = tensor<string, []>("op_311_cast_fp16")];
145 tensor<int32, [4]> var_318_begin_0 = const()[name = tensor<string, []>("op_318_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
146 tensor<int32, [4]> var_318_end_0 = const()[name = tensor<string, []>("op_318_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
147 tensor<bool, [4]> var_318_end_mask_0 = const()[name = tensor<string, []>("op_318_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
148 tensor<fp16, [1, 64, 1, 375]> var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, x = var_196_cast_fp16)[name = tensor<string, []>("op_318_cast_fp16")];
149 tensor<int32, [4]> var_325_begin_0 = const()[name = tensor<string, []>("op_325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
150 tensor<int32, [4]> var_325_end_0 = const()[name = tensor<string, []>("op_325_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
151 tensor<bool, [4]> var_325_end_mask_0 = const()[name = tensor<string, []>("op_325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
152 tensor<fp16, [1, 64, 1, 375]> var_325_cast_fp16 = slice_by_index(begin = var_325_begin_0, end = var_325_end_0, end_mask = var_325_end_mask_0, x = var_200_cast_fp16)[name = tensor<string, []>("op_325_cast_fp16")];
153 tensor<int32, [4]> var_332_begin_0 = const()[name = tensor<string, []>("op_332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
154 tensor<int32, [4]> var_332_end_0 = const()[name = tensor<string, []>("op_332_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
155 tensor<bool, [4]> var_332_end_mask_0 = const()[name = tensor<string, []>("op_332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
156 tensor<fp16, [1, 64, 1, 375]> var_332_cast_fp16 = slice_by_index(begin = var_332_begin_0, end = var_332_end_0, end_mask = var_332_end_mask_0, x = var_200_cast_fp16)[name = tensor<string, []>("op_332_cast_fp16")];
157 tensor<int32, [4]> var_339_begin_0 = const()[name = tensor<string, []>("op_339_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
158 tensor<int32, [4]> var_339_end_0 = const()[name = tensor<string, []>("op_339_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
159 tensor<bool, [4]> var_339_end_mask_0 = const()[name = tensor<string, []>("op_339_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
160 tensor<fp16, [1, 64, 1, 375]> var_339_cast_fp16 = slice_by_index(begin = var_339_begin_0, end = var_339_end_0, end_mask = var_339_end_mask_0, x = var_200_cast_fp16)[name = tensor<string, []>("op_339_cast_fp16")];
161 tensor<int32, [4]> var_346_begin_0 = const()[name = tensor<string, []>("op_346_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
162 tensor<int32, [4]> var_346_end_0 = const()[name = tensor<string, []>("op_346_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
163 tensor<bool, [4]> var_346_end_mask_0 = const()[name = tensor<string, []>("op_346_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
164 tensor<fp16, [1, 64, 1, 375]> var_346_cast_fp16 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, x = var_200_cast_fp16)[name = tensor<string, []>("op_346_cast_fp16")];
165 tensor<int32, [4]> var_353_begin_0 = const()[name = tensor<string, []>("op_353_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
166 tensor<int32, [4]> var_353_end_0 = const()[name = tensor<string, []>("op_353_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
167 tensor<bool, [4]> var_353_end_mask_0 = const()[name = tensor<string, []>("op_353_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
168 tensor<fp16, [1, 64, 1, 375]> var_353_cast_fp16 = slice_by_index(begin = var_353_begin_0, end = var_353_end_0, end_mask = var_353_end_mask_0, x = var_204_cast_fp16)[name = tensor<string, []>("op_353_cast_fp16")];
169 tensor<int32, [4]> var_360_begin_0 = const()[name = tensor<string, []>("op_360_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
170 tensor<int32, [4]> var_360_end_0 = const()[name = tensor<string, []>("op_360_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
171 tensor<bool, [4]> var_360_end_mask_0 = const()[name = tensor<string, []>("op_360_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
172 tensor<fp16, [1, 64, 1, 375]> var_360_cast_fp16 = slice_by_index(begin = var_360_begin_0, end = var_360_end_0, end_mask = var_360_end_mask_0, x = var_204_cast_fp16)[name = tensor<string, []>("op_360_cast_fp16")];
173 tensor<int32, [4]> var_367_begin_0 = const()[name = tensor<string, []>("op_367_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
174 tensor<int32, [4]> var_367_end_0 = const()[name = tensor<string, []>("op_367_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
175 tensor<bool, [4]> var_367_end_mask_0 = const()[name = tensor<string, []>("op_367_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
176 tensor<fp16, [1, 64, 1, 375]> var_367_cast_fp16 = slice_by_index(begin = var_367_begin_0, end = var_367_end_0, end_mask = var_367_end_mask_0, x = var_204_cast_fp16)[name = tensor<string, []>("op_367_cast_fp16")];
177 tensor<int32, [4]> var_374_begin_0 = const()[name = tensor<string, []>("op_374_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
178 tensor<int32, [4]> var_374_end_0 = const()[name = tensor<string, []>("op_374_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
179 tensor<bool, [4]> var_374_end_mask_0 = const()[name = tensor<string, []>("op_374_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
180 tensor<fp16, [1, 64, 1, 375]> var_374_cast_fp16 = slice_by_index(begin = var_374_begin_0, end = var_374_end_0, end_mask = var_374_end_mask_0, x = var_204_cast_fp16)[name = tensor<string, []>("op_374_cast_fp16")];
181 tensor<int32, [4]> k_1_perm_0 = const()[name = tensor<string, []>("k_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
182 tensor<int32, [4]> var_379_begin_0 = const()[name = tensor<string, []>("op_379_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
183 tensor<int32, [4]> var_379_end_0 = const()[name = tensor<string, []>("op_379_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 64])];
184 tensor<bool, [4]> var_379_end_mask_0 = const()[name = tensor<string, []>("op_379_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
185 tensor<fp16, [1, 1500, 1, 384]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = key_1_cast_fp16)[name = tensor<string, []>("transpose_3")];
186 tensor<fp16, [1, 1500, 1, 64]> var_379_cast_fp16 = slice_by_index(begin = var_379_begin_0, end = var_379_end_0, end_mask = var_379_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("op_379_cast_fp16")];
187 tensor<int32, [4]> var_383_begin_0 = const()[name = tensor<string, []>("op_383_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
188 tensor<int32, [4]> var_383_end_0 = const()[name = tensor<string, []>("op_383_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 128])];
189 tensor<bool, [4]> var_383_end_mask_0 = const()[name = tensor<string, []>("op_383_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
190 tensor<fp16, [1, 1500, 1, 64]> var_383_cast_fp16 = slice_by_index(begin = var_383_begin_0, end = var_383_end_0, end_mask = var_383_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("op_383_cast_fp16")];
191 tensor<int32, [4]> var_387_begin_0 = const()[name = tensor<string, []>("op_387_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
192 tensor<int32, [4]> var_387_end_0 = const()[name = tensor<string, []>("op_387_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 192])];
193 tensor<bool, [4]> var_387_end_mask_0 = const()[name = tensor<string, []>("op_387_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
194 tensor<fp16, [1, 1500, 1, 64]> var_387_cast_fp16 = slice_by_index(begin = var_387_begin_0, end = var_387_end_0, end_mask = var_387_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("op_387_cast_fp16")];
195 tensor<int32, [4]> var_391_begin_0 = const()[name = tensor<string, []>("op_391_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 192])];
196 tensor<int32, [4]> var_391_end_0 = const()[name = tensor<string, []>("op_391_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 256])];
197 tensor<bool, [4]> var_391_end_mask_0 = const()[name = tensor<string, []>("op_391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
198 tensor<fp16, [1, 1500, 1, 64]> var_391_cast_fp16 = slice_by_index(begin = var_391_begin_0, end = var_391_end_0, end_mask = var_391_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("op_391_cast_fp16")];
199 tensor<int32, [4]> var_395_begin_0 = const()[name = tensor<string, []>("op_395_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
200 tensor<int32, [4]> var_395_end_0 = const()[name = tensor<string, []>("op_395_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 320])];
201 tensor<bool, [4]> var_395_end_mask_0 = const()[name = tensor<string, []>("op_395_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
202 tensor<fp16, [1, 1500, 1, 64]> var_395_cast_fp16 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("op_395_cast_fp16")];
203 tensor<int32, [4]> var_399_begin_0 = const()[name = tensor<string, []>("op_399_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 320])];
204 tensor<int32, [4]> var_399_end_0 = const()[name = tensor<string, []>("op_399_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 384])];
205 tensor<bool, [4]> var_399_end_mask_0 = const()[name = tensor<string, []>("op_399_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
206 tensor<fp16, [1, 1500, 1, 64]> var_399_cast_fp16 = slice_by_index(begin = var_399_begin_0, end = var_399_end_0, end_mask = var_399_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("op_399_cast_fp16")];
207 tensor<int32, [4]> var_401_begin_0 = const()[name = tensor<string, []>("op_401_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
208 tensor<int32, [4]> var_401_end_0 = const()[name = tensor<string, []>("op_401_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
209 tensor<bool, [4]> var_401_end_mask_0 = const()[name = tensor<string, []>("op_401_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
210 tensor<fp16, [1, 64, 1, 1500]> var_401_cast_fp16 = slice_by_index(begin = var_401_begin_0, end = var_401_end_0, end_mask = var_401_end_mask_0, x = value_1_cast_fp16)[name = tensor<string, []>("op_401_cast_fp16")];
211 tensor<int32, [4]> var_405_begin_0 = const()[name = tensor<string, []>("op_405_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
212 tensor<int32, [4]> var_405_end_0 = const()[name = tensor<string, []>("op_405_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
213 tensor<bool, [4]> var_405_end_mask_0 = const()[name = tensor<string, []>("op_405_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
214 tensor<fp16, [1, 64, 1, 1500]> var_405_cast_fp16 = slice_by_index(begin = var_405_begin_0, end = var_405_end_0, end_mask = var_405_end_mask_0, x = value_1_cast_fp16)[name = tensor<string, []>("op_405_cast_fp16")];
215 tensor<int32, [4]> var_409_begin_0 = const()[name = tensor<string, []>("op_409_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
216 tensor<int32, [4]> var_409_end_0 = const()[name = tensor<string, []>("op_409_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
217 tensor<bool, [4]> var_409_end_mask_0 = const()[name = tensor<string, []>("op_409_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
218 tensor<fp16, [1, 64, 1, 1500]> var_409_cast_fp16 = slice_by_index(begin = var_409_begin_0, end = var_409_end_0, end_mask = var_409_end_mask_0, x = value_1_cast_fp16)[name = tensor<string, []>("op_409_cast_fp16")];
219 tensor<int32, [4]> var_413_begin_0 = const()[name = tensor<string, []>("op_413_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
220 tensor<int32, [4]> var_413_end_0 = const()[name = tensor<string, []>("op_413_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
221 tensor<bool, [4]> var_413_end_mask_0 = const()[name = tensor<string, []>("op_413_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
222 tensor<fp16, [1, 64, 1, 1500]> var_413_cast_fp16 = slice_by_index(begin = var_413_begin_0, end = var_413_end_0, end_mask = var_413_end_mask_0, x = value_1_cast_fp16)[name = tensor<string, []>("op_413_cast_fp16")];
223 tensor<int32, [4]> var_417_begin_0 = const()[name = tensor<string, []>("op_417_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
224 tensor<int32, [4]> var_417_end_0 = const()[name = tensor<string, []>("op_417_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
225 tensor<bool, [4]> var_417_end_mask_0 = const()[name = tensor<string, []>("op_417_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
226 tensor<fp16, [1, 64, 1, 1500]> var_417_cast_fp16 = slice_by_index(begin = var_417_begin_0, end = var_417_end_0, end_mask = var_417_end_mask_0, x = value_1_cast_fp16)[name = tensor<string, []>("op_417_cast_fp16")];
227 tensor<int32, [4]> var_421_begin_0 = const()[name = tensor<string, []>("op_421_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
228 tensor<int32, [4]> var_421_end_0 = const()[name = tensor<string, []>("op_421_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
229 tensor<bool, [4]> var_421_end_mask_0 = const()[name = tensor<string, []>("op_421_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
230 tensor<fp16, [1, 64, 1, 1500]> var_421_cast_fp16 = slice_by_index(begin = var_421_begin_0, end = var_421_end_0, end_mask = var_421_end_mask_0, x = value_1_cast_fp16)[name = tensor<string, []>("op_421_cast_fp16")];
231 tensor<string, []> _SplitHeadsQ__mh_w_1_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_1_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
232 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_1_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_1_equation_0, values = (var_379_cast_fp16, var_213_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_1_cast_fp16")];
233 tensor<string, []> _SplitHeadsQ__mh_w_3_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_3_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
234 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_3_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_3_equation_0, values = (var_379_cast_fp16, var_220_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_3_cast_fp16")];
235 tensor<string, []> _SplitHeadsQ__mh_w_5_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_5_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
236 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_5_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_5_equation_0, values = (var_379_cast_fp16, var_227_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_5_cast_fp16")];
237 tensor<string, []> _SplitHeadsQ__mh_w_7_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_7_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
238 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_7_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_7_equation_0, values = (var_379_cast_fp16, var_234_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_7_cast_fp16")];
239 tensor<string, []> _SplitHeadsQ__mh_w_9_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_9_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
240 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_9_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_9_equation_0, values = (var_383_cast_fp16, var_241_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_9_cast_fp16")];
241 tensor<string, []> _SplitHeadsQ__mh_w_11_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_11_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
242 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_11_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_11_equation_0, values = (var_383_cast_fp16, var_248_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_11_cast_fp16")];
243 tensor<string, []> _SplitHeadsQ__mh_w_13_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_13_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
244 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_13_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_13_equation_0, values = (var_383_cast_fp16, var_255_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_13_cast_fp16")];
245 tensor<string, []> _SplitHeadsQ__mh_w_15_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_15_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
246 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_15_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_15_equation_0, values = (var_383_cast_fp16, var_262_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_15_cast_fp16")];
247 tensor<string, []> _SplitHeadsQ__mh_w_17_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_17_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
248 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_17_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_17_equation_0, values = (var_387_cast_fp16, var_269_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_17_cast_fp16")];
249 tensor<string, []> _SplitHeadsQ__mh_w_19_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_19_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
250 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_19_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_19_equation_0, values = (var_387_cast_fp16, var_276_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_19_cast_fp16")];
251 tensor<string, []> _SplitHeadsQ__mh_w_21_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_21_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
252 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_21_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_21_equation_0, values = (var_387_cast_fp16, var_283_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_21_cast_fp16")];
253 tensor<string, []> _SplitHeadsQ__mh_w_23_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_23_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
254 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_23_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_23_equation_0, values = (var_387_cast_fp16, var_290_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_23_cast_fp16")];
255 tensor<string, []> _SplitHeadsQ__mh_w_25_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_25_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
256 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_25_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_25_equation_0, values = (var_391_cast_fp16, var_297_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_25_cast_fp16")];
257 tensor<string, []> _SplitHeadsQ__mh_w_27_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_27_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
258 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_27_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_27_equation_0, values = (var_391_cast_fp16, var_304_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_27_cast_fp16")];
259 tensor<string, []> _SplitHeadsQ__mh_w_29_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_29_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
260 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_29_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_29_equation_0, values = (var_391_cast_fp16, var_311_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_29_cast_fp16")];
261 tensor<string, []> _SplitHeadsQ__mh_w_31_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_31_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
262 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_31_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_31_equation_0, values = (var_391_cast_fp16, var_318_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_31_cast_fp16")];
263 tensor<string, []> _SplitHeadsQ__mh_w_33_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_33_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
264 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_33_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_33_equation_0, values = (var_395_cast_fp16, var_325_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_33_cast_fp16")];
265 tensor<string, []> _SplitHeadsQ__mh_w_35_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_35_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
266 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_35_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_35_equation_0, values = (var_395_cast_fp16, var_332_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_35_cast_fp16")];
267 tensor<string, []> _SplitHeadsQ__mh_w_37_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_37_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
268 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_37_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_37_equation_0, values = (var_395_cast_fp16, var_339_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_37_cast_fp16")];
269 tensor<string, []> _SplitHeadsQ__mh_w_39_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_39_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
270 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_39_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_39_equation_0, values = (var_395_cast_fp16, var_346_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_39_cast_fp16")];
271 tensor<string, []> _SplitHeadsQ__mh_w_41_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_41_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
272 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_41_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_41_equation_0, values = (var_399_cast_fp16, var_353_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_41_cast_fp16")];
273 tensor<string, []> _SplitHeadsQ__mh_w_43_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_43_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
274 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_43_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_43_equation_0, values = (var_399_cast_fp16, var_360_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_43_cast_fp16")];
275 tensor<string, []> _SplitHeadsQ__mh_w_45_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_45_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
276 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_45_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_45_equation_0, values = (var_399_cast_fp16, var_367_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_45_cast_fp16")];
277 tensor<string, []> _SplitHeadsQ__mh_w_47_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_47_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
278 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_47_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_47_equation_0, values = (var_399_cast_fp16, var_374_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_47_cast_fp16")];
279 tensor<fp16, []> var_472_to_fp16 = const()[name = tensor<string, []>("op_472_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
280 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_1_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_1_cast_fp16, y = var_472_to_fp16)[name = tensor<string, []>("aw_chunk_1_cast_fp16")];
281 tensor<fp16, []> var_474_to_fp16 = const()[name = tensor<string, []>("op_474_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
282 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_3_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_3_cast_fp16, y = var_474_to_fp16)[name = tensor<string, []>("aw_chunk_3_cast_fp16")];
283 tensor<fp16, []> var_476_to_fp16 = const()[name = tensor<string, []>("op_476_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
284 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_5_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_5_cast_fp16, y = var_476_to_fp16)[name = tensor<string, []>("aw_chunk_5_cast_fp16")];
285 tensor<fp16, []> var_478_to_fp16 = const()[name = tensor<string, []>("op_478_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
286 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_7_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_7_cast_fp16, y = var_478_to_fp16)[name = tensor<string, []>("aw_chunk_7_cast_fp16")];
287 tensor<fp16, []> var_480_to_fp16 = const()[name = tensor<string, []>("op_480_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
288 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_9_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_9_cast_fp16, y = var_480_to_fp16)[name = tensor<string, []>("aw_chunk_9_cast_fp16")];
289 tensor<fp16, []> var_482_to_fp16 = const()[name = tensor<string, []>("op_482_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
290 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_11_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_11_cast_fp16, y = var_482_to_fp16)[name = tensor<string, []>("aw_chunk_11_cast_fp16")];
291 tensor<fp16, []> var_484_to_fp16 = const()[name = tensor<string, []>("op_484_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
292 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_13_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_13_cast_fp16, y = var_484_to_fp16)[name = tensor<string, []>("aw_chunk_13_cast_fp16")];
293 tensor<fp16, []> var_486_to_fp16 = const()[name = tensor<string, []>("op_486_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
294 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_15_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_15_cast_fp16, y = var_486_to_fp16)[name = tensor<string, []>("aw_chunk_15_cast_fp16")];
295 tensor<fp16, []> var_488_to_fp16 = const()[name = tensor<string, []>("op_488_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
296 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_17_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_17_cast_fp16, y = var_488_to_fp16)[name = tensor<string, []>("aw_chunk_17_cast_fp16")];
297 tensor<fp16, []> var_490_to_fp16 = const()[name = tensor<string, []>("op_490_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
298 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_19_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_19_cast_fp16, y = var_490_to_fp16)[name = tensor<string, []>("aw_chunk_19_cast_fp16")];
299 tensor<fp16, []> var_492_to_fp16 = const()[name = tensor<string, []>("op_492_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
300 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_21_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_21_cast_fp16, y = var_492_to_fp16)[name = tensor<string, []>("aw_chunk_21_cast_fp16")];
301 tensor<fp16, []> var_494_to_fp16 = const()[name = tensor<string, []>("op_494_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
302 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_23_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_23_cast_fp16, y = var_494_to_fp16)[name = tensor<string, []>("aw_chunk_23_cast_fp16")];
303 tensor<fp16, []> var_496_to_fp16 = const()[name = tensor<string, []>("op_496_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
304 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_25_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_25_cast_fp16, y = var_496_to_fp16)[name = tensor<string, []>("aw_chunk_25_cast_fp16")];
305 tensor<fp16, []> var_498_to_fp16 = const()[name = tensor<string, []>("op_498_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
306 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_27_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_27_cast_fp16, y = var_498_to_fp16)[name = tensor<string, []>("aw_chunk_27_cast_fp16")];
307 tensor<fp16, []> var_500_to_fp16 = const()[name = tensor<string, []>("op_500_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
308 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_29_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_29_cast_fp16, y = var_500_to_fp16)[name = tensor<string, []>("aw_chunk_29_cast_fp16")];
309 tensor<fp16, []> var_502_to_fp16 = const()[name = tensor<string, []>("op_502_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
310 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_31_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_31_cast_fp16, y = var_502_to_fp16)[name = tensor<string, []>("aw_chunk_31_cast_fp16")];
311 tensor<fp16, []> var_504_to_fp16 = const()[name = tensor<string, []>("op_504_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
312 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_33_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_33_cast_fp16, y = var_504_to_fp16)[name = tensor<string, []>("aw_chunk_33_cast_fp16")];
313 tensor<fp16, []> var_506_to_fp16 = const()[name = tensor<string, []>("op_506_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
314 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_35_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_35_cast_fp16, y = var_506_to_fp16)[name = tensor<string, []>("aw_chunk_35_cast_fp16")];
315 tensor<fp16, []> var_508_to_fp16 = const()[name = tensor<string, []>("op_508_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
316 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_37_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_37_cast_fp16, y = var_508_to_fp16)[name = tensor<string, []>("aw_chunk_37_cast_fp16")];
317 tensor<fp16, []> var_510_to_fp16 = const()[name = tensor<string, []>("op_510_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
318 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_39_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_39_cast_fp16, y = var_510_to_fp16)[name = tensor<string, []>("aw_chunk_39_cast_fp16")];
319 tensor<fp16, []> var_512_to_fp16 = const()[name = tensor<string, []>("op_512_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
320 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_41_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_41_cast_fp16, y = var_512_to_fp16)[name = tensor<string, []>("aw_chunk_41_cast_fp16")];
321 tensor<fp16, []> var_514_to_fp16 = const()[name = tensor<string, []>("op_514_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
322 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_43_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_43_cast_fp16, y = var_514_to_fp16)[name = tensor<string, []>("aw_chunk_43_cast_fp16")];
323 tensor<fp16, []> var_516_to_fp16 = const()[name = tensor<string, []>("op_516_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
324 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_45_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_45_cast_fp16, y = var_516_to_fp16)[name = tensor<string, []>("aw_chunk_45_cast_fp16")];
325 tensor<fp16, []> var_518_to_fp16 = const()[name = tensor<string, []>("op_518_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
326 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_47_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_47_cast_fp16, y = var_518_to_fp16)[name = tensor<string, []>("aw_chunk_47_cast_fp16")];
327 tensor<fp16, [1, 1500, 1, 375]> var_520_cast_fp16 = softmax(axis = var_129, x = aw_chunk_1_cast_fp16)[name = tensor<string, []>("op_520_cast_fp16")];
328 tensor<fp16, [1, 1500, 1, 375]> var_521_cast_fp16 = softmax(axis = var_129, x = aw_chunk_3_cast_fp16)[name = tensor<string, []>("op_521_cast_fp16")];
329 tensor<fp16, [1, 1500, 1, 375]> var_522_cast_fp16 = softmax(axis = var_129, x = aw_chunk_5_cast_fp16)[name = tensor<string, []>("op_522_cast_fp16")];
330 tensor<fp16, [1, 1500, 1, 375]> var_523_cast_fp16 = softmax(axis = var_129, x = aw_chunk_7_cast_fp16)[name = tensor<string, []>("op_523_cast_fp16")];
331 tensor<fp16, [1, 1500, 1, 375]> var_524_cast_fp16 = softmax(axis = var_129, x = aw_chunk_9_cast_fp16)[name = tensor<string, []>("op_524_cast_fp16")];
332 tensor<fp16, [1, 1500, 1, 375]> var_525_cast_fp16 = softmax(axis = var_129, x = aw_chunk_11_cast_fp16)[name = tensor<string, []>("op_525_cast_fp16")];
333 tensor<fp16, [1, 1500, 1, 375]> var_526_cast_fp16 = softmax(axis = var_129, x = aw_chunk_13_cast_fp16)[name = tensor<string, []>("op_526_cast_fp16")];
334 tensor<fp16, [1, 1500, 1, 375]> var_527_cast_fp16 = softmax(axis = var_129, x = aw_chunk_15_cast_fp16)[name = tensor<string, []>("op_527_cast_fp16")];
335 tensor<fp16, [1, 1500, 1, 375]> var_528_cast_fp16 = softmax(axis = var_129, x = aw_chunk_17_cast_fp16)[name = tensor<string, []>("op_528_cast_fp16")];
336 tensor<fp16, [1, 1500, 1, 375]> var_529_cast_fp16 = softmax(axis = var_129, x = aw_chunk_19_cast_fp16)[name = tensor<string, []>("op_529_cast_fp16")];
337 tensor<fp16, [1, 1500, 1, 375]> var_530_cast_fp16 = softmax(axis = var_129, x = aw_chunk_21_cast_fp16)[name = tensor<string, []>("op_530_cast_fp16")];
338 tensor<fp16, [1, 1500, 1, 375]> var_531_cast_fp16 = softmax(axis = var_129, x = aw_chunk_23_cast_fp16)[name = tensor<string, []>("op_531_cast_fp16")];
339 tensor<fp16, [1, 1500, 1, 375]> var_532_cast_fp16 = softmax(axis = var_129, x = aw_chunk_25_cast_fp16)[name = tensor<string, []>("op_532_cast_fp16")];
340 tensor<fp16, [1, 1500, 1, 375]> var_533_cast_fp16 = softmax(axis = var_129, x = aw_chunk_27_cast_fp16)[name = tensor<string, []>("op_533_cast_fp16")];
341 tensor<fp16, [1, 1500, 1, 375]> var_534_cast_fp16 = softmax(axis = var_129, x = aw_chunk_29_cast_fp16)[name = tensor<string, []>("op_534_cast_fp16")];
342 tensor<fp16, [1, 1500, 1, 375]> var_535_cast_fp16 = softmax(axis = var_129, x = aw_chunk_31_cast_fp16)[name = tensor<string, []>("op_535_cast_fp16")];
343 tensor<fp16, [1, 1500, 1, 375]> var_536_cast_fp16 = softmax(axis = var_129, x = aw_chunk_33_cast_fp16)[name = tensor<string, []>("op_536_cast_fp16")];
344 tensor<fp16, [1, 1500, 1, 375]> var_537_cast_fp16 = softmax(axis = var_129, x = aw_chunk_35_cast_fp16)[name = tensor<string, []>("op_537_cast_fp16")];
345 tensor<fp16, [1, 1500, 1, 375]> var_538_cast_fp16 = softmax(axis = var_129, x = aw_chunk_37_cast_fp16)[name = tensor<string, []>("op_538_cast_fp16")];
346 tensor<fp16, [1, 1500, 1, 375]> var_539_cast_fp16 = softmax(axis = var_129, x = aw_chunk_39_cast_fp16)[name = tensor<string, []>("op_539_cast_fp16")];
347 tensor<fp16, [1, 1500, 1, 375]> var_540_cast_fp16 = softmax(axis = var_129, x = aw_chunk_41_cast_fp16)[name = tensor<string, []>("op_540_cast_fp16")];
348 tensor<fp16, [1, 1500, 1, 375]> var_541_cast_fp16 = softmax(axis = var_129, x = aw_chunk_43_cast_fp16)[name = tensor<string, []>("op_541_cast_fp16")];
349 tensor<fp16, [1, 1500, 1, 375]> var_542_cast_fp16 = softmax(axis = var_129, x = aw_chunk_45_cast_fp16)[name = tensor<string, []>("op_542_cast_fp16")];
350 tensor<fp16, [1, 1500, 1, 375]> var_543_cast_fp16 = softmax(axis = var_129, x = aw_chunk_47_cast_fp16)[name = tensor<string, []>("op_543_cast_fp16")];
351 tensor<string, []> var_545_equation_0 = const()[name = tensor<string, []>("op_545_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
352 tensor<fp16, [1, 64, 1, 375]> var_545_cast_fp16 = einsum(equation = var_545_equation_0, values = (var_401_cast_fp16, var_520_cast_fp16))[name = tensor<string, []>("op_545_cast_fp16")];
353 tensor<string, []> var_547_equation_0 = const()[name = tensor<string, []>("op_547_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
354 tensor<fp16, [1, 64, 1, 375]> var_547_cast_fp16 = einsum(equation = var_547_equation_0, values = (var_401_cast_fp16, var_521_cast_fp16))[name = tensor<string, []>("op_547_cast_fp16")];
355 tensor<string, []> var_549_equation_0 = const()[name = tensor<string, []>("op_549_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
356 tensor<fp16, [1, 64, 1, 375]> var_549_cast_fp16 = einsum(equation = var_549_equation_0, values = (var_401_cast_fp16, var_522_cast_fp16))[name = tensor<string, []>("op_549_cast_fp16")];
357 tensor<string, []> var_551_equation_0 = const()[name = tensor<string, []>("op_551_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
358 tensor<fp16, [1, 64, 1, 375]> var_551_cast_fp16 = einsum(equation = var_551_equation_0, values = (var_401_cast_fp16, var_523_cast_fp16))[name = tensor<string, []>("op_551_cast_fp16")];
359 tensor<string, []> var_553_equation_0 = const()[name = tensor<string, []>("op_553_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
360 tensor<fp16, [1, 64, 1, 375]> var_553_cast_fp16 = einsum(equation = var_553_equation_0, values = (var_405_cast_fp16, var_524_cast_fp16))[name = tensor<string, []>("op_553_cast_fp16")];
361 tensor<string, []> var_555_equation_0 = const()[name = tensor<string, []>("op_555_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
362 tensor<fp16, [1, 64, 1, 375]> var_555_cast_fp16 = einsum(equation = var_555_equation_0, values = (var_405_cast_fp16, var_525_cast_fp16))[name = tensor<string, []>("op_555_cast_fp16")];
363 tensor<string, []> var_557_equation_0 = const()[name = tensor<string, []>("op_557_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
364 tensor<fp16, [1, 64, 1, 375]> var_557_cast_fp16 = einsum(equation = var_557_equation_0, values = (var_405_cast_fp16, var_526_cast_fp16))[name = tensor<string, []>("op_557_cast_fp16")];
365 tensor<string, []> var_559_equation_0 = const()[name = tensor<string, []>("op_559_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
366 tensor<fp16, [1, 64, 1, 375]> var_559_cast_fp16 = einsum(equation = var_559_equation_0, values = (var_405_cast_fp16, var_527_cast_fp16))[name = tensor<string, []>("op_559_cast_fp16")];
367 tensor<string, []> var_561_equation_0 = const()[name = tensor<string, []>("op_561_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
368 tensor<fp16, [1, 64, 1, 375]> var_561_cast_fp16 = einsum(equation = var_561_equation_0, values = (var_409_cast_fp16, var_528_cast_fp16))[name = tensor<string, []>("op_561_cast_fp16")];
369 tensor<string, []> var_563_equation_0 = const()[name = tensor<string, []>("op_563_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
370 tensor<fp16, [1, 64, 1, 375]> var_563_cast_fp16 = einsum(equation = var_563_equation_0, values = (var_409_cast_fp16, var_529_cast_fp16))[name = tensor<string, []>("op_563_cast_fp16")];
371 tensor<string, []> var_565_equation_0 = const()[name = tensor<string, []>("op_565_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
372 tensor<fp16, [1, 64, 1, 375]> var_565_cast_fp16 = einsum(equation = var_565_equation_0, values = (var_409_cast_fp16, var_530_cast_fp16))[name = tensor<string, []>("op_565_cast_fp16")];
373 tensor<string, []> var_567_equation_0 = const()[name = tensor<string, []>("op_567_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
374 tensor<fp16, [1, 64, 1, 375]> var_567_cast_fp16 = einsum(equation = var_567_equation_0, values = (var_409_cast_fp16, var_531_cast_fp16))[name = tensor<string, []>("op_567_cast_fp16")];
375 tensor<string, []> var_569_equation_0 = const()[name = tensor<string, []>("op_569_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
376 tensor<fp16, [1, 64, 1, 375]> var_569_cast_fp16 = einsum(equation = var_569_equation_0, values = (var_413_cast_fp16, var_532_cast_fp16))[name = tensor<string, []>("op_569_cast_fp16")];
377 tensor<string, []> var_571_equation_0 = const()[name = tensor<string, []>("op_571_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
378 tensor<fp16, [1, 64, 1, 375]> var_571_cast_fp16 = einsum(equation = var_571_equation_0, values = (var_413_cast_fp16, var_533_cast_fp16))[name = tensor<string, []>("op_571_cast_fp16")];
379 tensor<string, []> var_573_equation_0 = const()[name = tensor<string, []>("op_573_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
380 tensor<fp16, [1, 64, 1, 375]> var_573_cast_fp16 = einsum(equation = var_573_equation_0, values = (var_413_cast_fp16, var_534_cast_fp16))[name = tensor<string, []>("op_573_cast_fp16")];
381 tensor<string, []> var_575_equation_0 = const()[name = tensor<string, []>("op_575_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
382 tensor<fp16, [1, 64, 1, 375]> var_575_cast_fp16 = einsum(equation = var_575_equation_0, values = (var_413_cast_fp16, var_535_cast_fp16))[name = tensor<string, []>("op_575_cast_fp16")];
383 tensor<string, []> var_577_equation_0 = const()[name = tensor<string, []>("op_577_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
384 tensor<fp16, [1, 64, 1, 375]> var_577_cast_fp16 = einsum(equation = var_577_equation_0, values = (var_417_cast_fp16, var_536_cast_fp16))[name = tensor<string, []>("op_577_cast_fp16")];
385 tensor<string, []> var_579_equation_0 = const()[name = tensor<string, []>("op_579_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
386 tensor<fp16, [1, 64, 1, 375]> var_579_cast_fp16 = einsum(equation = var_579_equation_0, values = (var_417_cast_fp16, var_537_cast_fp16))[name = tensor<string, []>("op_579_cast_fp16")];
387 tensor<string, []> var_581_equation_0 = const()[name = tensor<string, []>("op_581_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
388 tensor<fp16, [1, 64, 1, 375]> var_581_cast_fp16 = einsum(equation = var_581_equation_0, values = (var_417_cast_fp16, var_538_cast_fp16))[name = tensor<string, []>("op_581_cast_fp16")];
389 tensor<string, []> var_583_equation_0 = const()[name = tensor<string, []>("op_583_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
390 tensor<fp16, [1, 64, 1, 375]> var_583_cast_fp16 = einsum(equation = var_583_equation_0, values = (var_417_cast_fp16, var_539_cast_fp16))[name = tensor<string, []>("op_583_cast_fp16")];
391 tensor<string, []> var_585_equation_0 = const()[name = tensor<string, []>("op_585_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
392 tensor<fp16, [1, 64, 1, 375]> var_585_cast_fp16 = einsum(equation = var_585_equation_0, values = (var_421_cast_fp16, var_540_cast_fp16))[name = tensor<string, []>("op_585_cast_fp16")];
393 tensor<string, []> var_587_equation_0 = const()[name = tensor<string, []>("op_587_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
394 tensor<fp16, [1, 64, 1, 375]> var_587_cast_fp16 = einsum(equation = var_587_equation_0, values = (var_421_cast_fp16, var_541_cast_fp16))[name = tensor<string, []>("op_587_cast_fp16")];
395 tensor<string, []> var_589_equation_0 = const()[name = tensor<string, []>("op_589_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
396 tensor<fp16, [1, 64, 1, 375]> var_589_cast_fp16 = einsum(equation = var_589_equation_0, values = (var_421_cast_fp16, var_542_cast_fp16))[name = tensor<string, []>("op_589_cast_fp16")];
397 tensor<string, []> var_591_equation_0 = const()[name = tensor<string, []>("op_591_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
398 tensor<fp16, [1, 64, 1, 375]> var_591_cast_fp16 = einsum(equation = var_591_equation_0, values = (var_421_cast_fp16, var_543_cast_fp16))[name = tensor<string, []>("op_591_cast_fp16")];
399 tensor<bool, []> var_593_interleave_0 = const()[name = tensor<string, []>("op_593_interleave_0"), val = tensor<bool, []>(false)];
400 tensor<fp16, [1, 64, 1, 1500]> var_593_cast_fp16 = concat(axis = var_118, interleave = var_593_interleave_0, values = (var_545_cast_fp16, var_547_cast_fp16, var_549_cast_fp16, var_551_cast_fp16))[name = tensor<string, []>("op_593_cast_fp16")];
401 tensor<bool, []> var_595_interleave_0 = const()[name = tensor<string, []>("op_595_interleave_0"), val = tensor<bool, []>(false)];
402 tensor<fp16, [1, 64, 1, 1500]> var_595_cast_fp16 = concat(axis = var_118, interleave = var_595_interleave_0, values = (var_553_cast_fp16, var_555_cast_fp16, var_557_cast_fp16, var_559_cast_fp16))[name = tensor<string, []>("op_595_cast_fp16")];
403 tensor<bool, []> var_597_interleave_0 = const()[name = tensor<string, []>("op_597_interleave_0"), val = tensor<bool, []>(false)];
404 tensor<fp16, [1, 64, 1, 1500]> var_597_cast_fp16 = concat(axis = var_118, interleave = var_597_interleave_0, values = (var_561_cast_fp16, var_563_cast_fp16, var_565_cast_fp16, var_567_cast_fp16))[name = tensor<string, []>("op_597_cast_fp16")];
405 tensor<bool, []> var_599_interleave_0 = const()[name = tensor<string, []>("op_599_interleave_0"), val = tensor<bool, []>(false)];
406 tensor<fp16, [1, 64, 1, 1500]> var_599_cast_fp16 = concat(axis = var_118, interleave = var_599_interleave_0, values = (var_569_cast_fp16, var_571_cast_fp16, var_573_cast_fp16, var_575_cast_fp16))[name = tensor<string, []>("op_599_cast_fp16")];
407 tensor<bool, []> var_601_interleave_0 = const()[name = tensor<string, []>("op_601_interleave_0"), val = tensor<bool, []>(false)];
408 tensor<fp16, [1, 64, 1, 1500]> var_601_cast_fp16 = concat(axis = var_118, interleave = var_601_interleave_0, values = (var_577_cast_fp16, var_579_cast_fp16, var_581_cast_fp16, var_583_cast_fp16))[name = tensor<string, []>("op_601_cast_fp16")];
409 tensor<bool, []> var_603_interleave_0 = const()[name = tensor<string, []>("op_603_interleave_0"), val = tensor<bool, []>(false)];
410 tensor<fp16, [1, 64, 1, 1500]> var_603_cast_fp16 = concat(axis = var_118, interleave = var_603_interleave_0, values = (var_585_cast_fp16, var_587_cast_fp16, var_589_cast_fp16, var_591_cast_fp16))[name = tensor<string, []>("op_603_cast_fp16")];
411 tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
412 tensor<fp16, [1, 384, 1, 1500]> input_1_cast_fp16 = concat(axis = var_129, interleave = input_1_interleave_0, values = (var_593_cast_fp16, var_595_cast_fp16, var_597_cast_fp16, var_599_cast_fp16, var_601_cast_fp16, var_603_cast_fp16))[name = tensor<string, []>("input_1_cast_fp16")];
413 tensor<string, []> obj_3_pad_type_0 = const()[name = tensor<string, []>("obj_3_pad_type_0"), val = tensor<string, []>("valid")];
414 tensor<int32, [2]> obj_3_strides_0 = const()[name = tensor<string, []>("obj_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
415 tensor<int32, [4]> obj_3_pad_0 = const()[name = tensor<string, []>("obj_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
416 tensor<int32, [2]> obj_3_dilations_0 = const()[name = tensor<string, []>("obj_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
417 tensor<int32, []> obj_3_groups_0 = const()[name = tensor<string, []>("obj_3_groups_0"), val = tensor<int32, []>(1)];
418 tensor<fp16, [384, 384, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3112896)))];
419 tensor<fp16, [384]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407872)))];
420 tensor<fp16, [1, 384, 1, 1500]> obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_3_dilations_0, groups = obj_3_groups_0, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = obj_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_3_cast_fp16")];
421 tensor<fp16, [1, 384, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
422 tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
423 tensor<fp16, []> var_622_to_fp16 = const()[name = tensor<string, []>("op_622_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
424 tensor<fp16, [1, 384, 1, 1500]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_622_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
425 tensor<fp16, [384]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3408704)))];
426 tensor<fp16, [384]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3409536)))];
427 tensor<fp16, []> input_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
428 tensor<fp16, [1, 384, 1, 1500]> input_3_cast_fp16 = batch_norm(beta = input_3_beta_0_to_fp16, epsilon = input_3_epsilon_0_to_fp16, gamma = input_3_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, []>("input_3_cast_fp16")];
429 tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")];
430 tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
431 tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
432 tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
433 tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)];
434 tensor<fp16, [1536, 384, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3410368)))];
435 tensor<fp16, [1536]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4590080)))];
436 tensor<fp16, [1, 1536, 1, 1500]> input_5_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, 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_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
437 tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("EXACT")];
438 tensor<fp16, [1, 1536, 1, 1500]> input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
439 tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("valid")];
440 tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = tensor<string, []>("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
441 tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
442 tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = tensor<string, []>("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
443 tensor<int32, []> hidden_states_5_groups_0 = const()[name = tensor<string, []>("hidden_states_5_groups_0"), val = tensor<int32, []>(1)];
444 tensor<fp16, [384, 1536, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4593216)))];
445 tensor<fp16, [384]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5772928)))];
446 tensor<fp16, [1, 384, 1, 1500]> hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
447 tensor<fp16, [1, 384, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
448 tensor<int32, []> var_651 = const()[name = tensor<string, []>("op_651"), val = tensor<int32, []>(3)];
449 tensor<int32, []> var_662 = const()[name = tensor<string, []>("op_662"), val = tensor<int32, []>(1)];
450 tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
451 tensor<fp16, []> var_679_to_fp16 = const()[name = tensor<string, []>("op_679_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
452 tensor<fp16, [1, 384, 1, 1500]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_679_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
453 tensor<fp16, [384]> obj_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5773760)))];
454 tensor<fp16, [384]> obj_5_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_5_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5774592)))];
455 tensor<fp16, []> obj_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
456 tensor<fp16, [1, 384, 1, 1500]> obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_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, []>("obj_5_cast_fp16")];
457 tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("valid")];
458 tensor<int32, [2]> query_3_strides_0 = const()[name = tensor<string, []>("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
459 tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
460 tensor<int32, [2]> query_3_dilations_0 = const()[name = tensor<string, []>("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
461 tensor<int32, []> query_3_groups_0 = const()[name = tensor<string, []>("query_3_groups_0"), val = tensor<int32, []>(1)];
462 tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5775424)))];
463 tensor<fp16, [384]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6070400)))];
464 tensor<fp16, [1, 384, 1, 1500]> query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
465 tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")];
466 tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
467 tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
468 tensor<int32, [2]> key_3_dilations_0 = const()[name = tensor<string, []>("key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
469 tensor<int32, []> key_3_groups_0 = const()[name = tensor<string, []>("key_3_groups_0"), val = tensor<int32, []>(1)];
470 tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6071232)))];
471 tensor<fp16, [1, 384, 1, 1500]> key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
472 tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("valid")];
473 tensor<int32, [2]> value_3_strides_0 = const()[name = tensor<string, []>("value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
474 tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
475 tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
476 tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)];
477 tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6366208)))];
478 tensor<fp16, [384]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6661184)))];
479 tensor<fp16, [1, 384, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
480 tensor<int32, [4]> var_717_begin_0 = const()[name = tensor<string, []>("op_717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
481 tensor<int32, [4]> var_717_end_0 = const()[name = tensor<string, []>("op_717_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
482 tensor<bool, [4]> var_717_end_mask_0 = const()[name = tensor<string, []>("op_717_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
483 tensor<fp16, [1, 64, 1, 1500]> var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, x = query_3_cast_fp16)[name = tensor<string, []>("op_717_cast_fp16")];
484 tensor<int32, [4]> var_721_begin_0 = const()[name = tensor<string, []>("op_721_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
485 tensor<int32, [4]> var_721_end_0 = const()[name = tensor<string, []>("op_721_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
486 tensor<bool, [4]> var_721_end_mask_0 = const()[name = tensor<string, []>("op_721_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
487 tensor<fp16, [1, 64, 1, 1500]> var_721_cast_fp16 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = query_3_cast_fp16)[name = tensor<string, []>("op_721_cast_fp16")];
488 tensor<int32, [4]> var_725_begin_0 = const()[name = tensor<string, []>("op_725_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
489 tensor<int32, [4]> var_725_end_0 = const()[name = tensor<string, []>("op_725_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
490 tensor<bool, [4]> var_725_end_mask_0 = const()[name = tensor<string, []>("op_725_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
491 tensor<fp16, [1, 64, 1, 1500]> var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = query_3_cast_fp16)[name = tensor<string, []>("op_725_cast_fp16")];
492 tensor<int32, [4]> var_729_begin_0 = const()[name = tensor<string, []>("op_729_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
493 tensor<int32, [4]> var_729_end_0 = const()[name = tensor<string, []>("op_729_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
494 tensor<bool, [4]> var_729_end_mask_0 = const()[name = tensor<string, []>("op_729_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
495 tensor<fp16, [1, 64, 1, 1500]> var_729_cast_fp16 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = query_3_cast_fp16)[name = tensor<string, []>("op_729_cast_fp16")];
496 tensor<int32, [4]> var_733_begin_0 = const()[name = tensor<string, []>("op_733_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
497 tensor<int32, [4]> var_733_end_0 = const()[name = tensor<string, []>("op_733_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
498 tensor<bool, [4]> var_733_end_mask_0 = const()[name = tensor<string, []>("op_733_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
499 tensor<fp16, [1, 64, 1, 1500]> var_733_cast_fp16 = slice_by_index(begin = var_733_begin_0, end = var_733_end_0, end_mask = var_733_end_mask_0, x = query_3_cast_fp16)[name = tensor<string, []>("op_733_cast_fp16")];
500 tensor<int32, [4]> var_737_begin_0 = const()[name = tensor<string, []>("op_737_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
501 tensor<int32, [4]> var_737_end_0 = const()[name = tensor<string, []>("op_737_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
502 tensor<bool, [4]> var_737_end_mask_0 = const()[name = tensor<string, []>("op_737_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
503 tensor<fp16, [1, 64, 1, 1500]> var_737_cast_fp16 = slice_by_index(begin = var_737_begin_0, end = var_737_end_0, end_mask = var_737_end_mask_0, x = query_3_cast_fp16)[name = tensor<string, []>("op_737_cast_fp16")];
504 tensor<int32, [4]> var_746_begin_0 = const()[name = tensor<string, []>("op_746_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
505 tensor<int32, [4]> var_746_end_0 = const()[name = tensor<string, []>("op_746_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
506 tensor<bool, [4]> var_746_end_mask_0 = const()[name = tensor<string, []>("op_746_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
507 tensor<fp16, [1, 64, 1, 375]> var_746_cast_fp16 = slice_by_index(begin = var_746_begin_0, end = var_746_end_0, end_mask = var_746_end_mask_0, x = var_717_cast_fp16)[name = tensor<string, []>("op_746_cast_fp16")];
508 tensor<int32, [4]> var_753_begin_0 = const()[name = tensor<string, []>("op_753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
509 tensor<int32, [4]> var_753_end_0 = const()[name = tensor<string, []>("op_753_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
510 tensor<bool, [4]> var_753_end_mask_0 = const()[name = tensor<string, []>("op_753_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
511 tensor<fp16, [1, 64, 1, 375]> var_753_cast_fp16 = slice_by_index(begin = var_753_begin_0, end = var_753_end_0, end_mask = var_753_end_mask_0, x = var_717_cast_fp16)[name = tensor<string, []>("op_753_cast_fp16")];
512 tensor<int32, [4]> var_760_begin_0 = const()[name = tensor<string, []>("op_760_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
513 tensor<int32, [4]> var_760_end_0 = const()[name = tensor<string, []>("op_760_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
514 tensor<bool, [4]> var_760_end_mask_0 = const()[name = tensor<string, []>("op_760_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
515 tensor<fp16, [1, 64, 1, 375]> var_760_cast_fp16 = slice_by_index(begin = var_760_begin_0, end = var_760_end_0, end_mask = var_760_end_mask_0, x = var_717_cast_fp16)[name = tensor<string, []>("op_760_cast_fp16")];
516 tensor<int32, [4]> var_767_begin_0 = const()[name = tensor<string, []>("op_767_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
517 tensor<int32, [4]> var_767_end_0 = const()[name = tensor<string, []>("op_767_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
518 tensor<bool, [4]> var_767_end_mask_0 = const()[name = tensor<string, []>("op_767_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
519 tensor<fp16, [1, 64, 1, 375]> var_767_cast_fp16 = slice_by_index(begin = var_767_begin_0, end = var_767_end_0, end_mask = var_767_end_mask_0, x = var_717_cast_fp16)[name = tensor<string, []>("op_767_cast_fp16")];
520 tensor<int32, [4]> var_774_begin_0 = const()[name = tensor<string, []>("op_774_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
521 tensor<int32, [4]> var_774_end_0 = const()[name = tensor<string, []>("op_774_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
522 tensor<bool, [4]> var_774_end_mask_0 = const()[name = tensor<string, []>("op_774_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
523 tensor<fp16, [1, 64, 1, 375]> var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, x = var_721_cast_fp16)[name = tensor<string, []>("op_774_cast_fp16")];
524 tensor<int32, [4]> var_781_begin_0 = const()[name = tensor<string, []>("op_781_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
525 tensor<int32, [4]> var_781_end_0 = const()[name = tensor<string, []>("op_781_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
526 tensor<bool, [4]> var_781_end_mask_0 = const()[name = tensor<string, []>("op_781_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
527 tensor<fp16, [1, 64, 1, 375]> var_781_cast_fp16 = slice_by_index(begin = var_781_begin_0, end = var_781_end_0, end_mask = var_781_end_mask_0, x = var_721_cast_fp16)[name = tensor<string, []>("op_781_cast_fp16")];
528 tensor<int32, [4]> var_788_begin_0 = const()[name = tensor<string, []>("op_788_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
529 tensor<int32, [4]> var_788_end_0 = const()[name = tensor<string, []>("op_788_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
530 tensor<bool, [4]> var_788_end_mask_0 = const()[name = tensor<string, []>("op_788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
531 tensor<fp16, [1, 64, 1, 375]> var_788_cast_fp16 = slice_by_index(begin = var_788_begin_0, end = var_788_end_0, end_mask = var_788_end_mask_0, x = var_721_cast_fp16)[name = tensor<string, []>("op_788_cast_fp16")];
532 tensor<int32, [4]> var_795_begin_0 = const()[name = tensor<string, []>("op_795_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
533 tensor<int32, [4]> var_795_end_0 = const()[name = tensor<string, []>("op_795_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
534 tensor<bool, [4]> var_795_end_mask_0 = const()[name = tensor<string, []>("op_795_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
535 tensor<fp16, [1, 64, 1, 375]> var_795_cast_fp16 = slice_by_index(begin = var_795_begin_0, end = var_795_end_0, end_mask = var_795_end_mask_0, x = var_721_cast_fp16)[name = tensor<string, []>("op_795_cast_fp16")];
536 tensor<int32, [4]> var_802_begin_0 = const()[name = tensor<string, []>("op_802_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
537 tensor<int32, [4]> var_802_end_0 = const()[name = tensor<string, []>("op_802_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
538 tensor<bool, [4]> var_802_end_mask_0 = const()[name = tensor<string, []>("op_802_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
539 tensor<fp16, [1, 64, 1, 375]> var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, x = var_725_cast_fp16)[name = tensor<string, []>("op_802_cast_fp16")];
540 tensor<int32, [4]> var_809_begin_0 = const()[name = tensor<string, []>("op_809_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
541 tensor<int32, [4]> var_809_end_0 = const()[name = tensor<string, []>("op_809_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
542 tensor<bool, [4]> var_809_end_mask_0 = const()[name = tensor<string, []>("op_809_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
543 tensor<fp16, [1, 64, 1, 375]> var_809_cast_fp16 = slice_by_index(begin = var_809_begin_0, end = var_809_end_0, end_mask = var_809_end_mask_0, x = var_725_cast_fp16)[name = tensor<string, []>("op_809_cast_fp16")];
544 tensor<int32, [4]> var_816_begin_0 = const()[name = tensor<string, []>("op_816_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
545 tensor<int32, [4]> var_816_end_0 = const()[name = tensor<string, []>("op_816_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
546 tensor<bool, [4]> var_816_end_mask_0 = const()[name = tensor<string, []>("op_816_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
547 tensor<fp16, [1, 64, 1, 375]> var_816_cast_fp16 = slice_by_index(begin = var_816_begin_0, end = var_816_end_0, end_mask = var_816_end_mask_0, x = var_725_cast_fp16)[name = tensor<string, []>("op_816_cast_fp16")];
548 tensor<int32, [4]> var_823_begin_0 = const()[name = tensor<string, []>("op_823_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
549 tensor<int32, [4]> var_823_end_0 = const()[name = tensor<string, []>("op_823_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
550 tensor<bool, [4]> var_823_end_mask_0 = const()[name = tensor<string, []>("op_823_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
551 tensor<fp16, [1, 64, 1, 375]> var_823_cast_fp16 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = var_725_cast_fp16)[name = tensor<string, []>("op_823_cast_fp16")];
552 tensor<int32, [4]> var_830_begin_0 = const()[name = tensor<string, []>("op_830_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
553 tensor<int32, [4]> var_830_end_0 = const()[name = tensor<string, []>("op_830_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
554 tensor<bool, [4]> var_830_end_mask_0 = const()[name = tensor<string, []>("op_830_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
555 tensor<fp16, [1, 64, 1, 375]> var_830_cast_fp16 = slice_by_index(begin = var_830_begin_0, end = var_830_end_0, end_mask = var_830_end_mask_0, x = var_729_cast_fp16)[name = tensor<string, []>("op_830_cast_fp16")];
556 tensor<int32, [4]> var_837_begin_0 = const()[name = tensor<string, []>("op_837_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
557 tensor<int32, [4]> var_837_end_0 = const()[name = tensor<string, []>("op_837_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
558 tensor<bool, [4]> var_837_end_mask_0 = const()[name = tensor<string, []>("op_837_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
559 tensor<fp16, [1, 64, 1, 375]> var_837_cast_fp16 = slice_by_index(begin = var_837_begin_0, end = var_837_end_0, end_mask = var_837_end_mask_0, x = var_729_cast_fp16)[name = tensor<string, []>("op_837_cast_fp16")];
560 tensor<int32, [4]> var_844_begin_0 = const()[name = tensor<string, []>("op_844_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
561 tensor<int32, [4]> var_844_end_0 = const()[name = tensor<string, []>("op_844_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
562 tensor<bool, [4]> var_844_end_mask_0 = const()[name = tensor<string, []>("op_844_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
563 tensor<fp16, [1, 64, 1, 375]> var_844_cast_fp16 = slice_by_index(begin = var_844_begin_0, end = var_844_end_0, end_mask = var_844_end_mask_0, x = var_729_cast_fp16)[name = tensor<string, []>("op_844_cast_fp16")];
564 tensor<int32, [4]> var_851_begin_0 = const()[name = tensor<string, []>("op_851_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
565 tensor<int32, [4]> var_851_end_0 = const()[name = tensor<string, []>("op_851_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
566 tensor<bool, [4]> var_851_end_mask_0 = const()[name = tensor<string, []>("op_851_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
567 tensor<fp16, [1, 64, 1, 375]> var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = var_729_cast_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
568 tensor<int32, [4]> var_858_begin_0 = const()[name = tensor<string, []>("op_858_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
569 tensor<int32, [4]> var_858_end_0 = const()[name = tensor<string, []>("op_858_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
570 tensor<bool, [4]> var_858_end_mask_0 = const()[name = tensor<string, []>("op_858_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
571 tensor<fp16, [1, 64, 1, 375]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = var_733_cast_fp16)[name = tensor<string, []>("op_858_cast_fp16")];
572 tensor<int32, [4]> var_865_begin_0 = const()[name = tensor<string, []>("op_865_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
573 tensor<int32, [4]> var_865_end_0 = const()[name = tensor<string, []>("op_865_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
574 tensor<bool, [4]> var_865_end_mask_0 = const()[name = tensor<string, []>("op_865_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
575 tensor<fp16, [1, 64, 1, 375]> var_865_cast_fp16 = slice_by_index(begin = var_865_begin_0, end = var_865_end_0, end_mask = var_865_end_mask_0, x = var_733_cast_fp16)[name = tensor<string, []>("op_865_cast_fp16")];
576 tensor<int32, [4]> var_872_begin_0 = const()[name = tensor<string, []>("op_872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
577 tensor<int32, [4]> var_872_end_0 = const()[name = tensor<string, []>("op_872_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
578 tensor<bool, [4]> var_872_end_mask_0 = const()[name = tensor<string, []>("op_872_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
579 tensor<fp16, [1, 64, 1, 375]> var_872_cast_fp16 = slice_by_index(begin = var_872_begin_0, end = var_872_end_0, end_mask = var_872_end_mask_0, x = var_733_cast_fp16)[name = tensor<string, []>("op_872_cast_fp16")];
580 tensor<int32, [4]> var_879_begin_0 = const()[name = tensor<string, []>("op_879_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
581 tensor<int32, [4]> var_879_end_0 = const()[name = tensor<string, []>("op_879_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
582 tensor<bool, [4]> var_879_end_mask_0 = const()[name = tensor<string, []>("op_879_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
583 tensor<fp16, [1, 64, 1, 375]> var_879_cast_fp16 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = var_733_cast_fp16)[name = tensor<string, []>("op_879_cast_fp16")];
584 tensor<int32, [4]> var_886_begin_0 = const()[name = tensor<string, []>("op_886_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
585 tensor<int32, [4]> var_886_end_0 = const()[name = tensor<string, []>("op_886_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
586 tensor<bool, [4]> var_886_end_mask_0 = const()[name = tensor<string, []>("op_886_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
587 tensor<fp16, [1, 64, 1, 375]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = var_737_cast_fp16)[name = tensor<string, []>("op_886_cast_fp16")];
588 tensor<int32, [4]> var_893_begin_0 = const()[name = tensor<string, []>("op_893_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
589 tensor<int32, [4]> var_893_end_0 = const()[name = tensor<string, []>("op_893_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
590 tensor<bool, [4]> var_893_end_mask_0 = const()[name = tensor<string, []>("op_893_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
591 tensor<fp16, [1, 64, 1, 375]> var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = var_737_cast_fp16)[name = tensor<string, []>("op_893_cast_fp16")];
592 tensor<int32, [4]> var_900_begin_0 = const()[name = tensor<string, []>("op_900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
593 tensor<int32, [4]> var_900_end_0 = const()[name = tensor<string, []>("op_900_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
594 tensor<bool, [4]> var_900_end_mask_0 = const()[name = tensor<string, []>("op_900_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
595 tensor<fp16, [1, 64, 1, 375]> var_900_cast_fp16 = slice_by_index(begin = var_900_begin_0, end = var_900_end_0, end_mask = var_900_end_mask_0, x = var_737_cast_fp16)[name = tensor<string, []>("op_900_cast_fp16")];
596 tensor<int32, [4]> var_907_begin_0 = const()[name = tensor<string, []>("op_907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
597 tensor<int32, [4]> var_907_end_0 = const()[name = tensor<string, []>("op_907_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
598 tensor<bool, [4]> var_907_end_mask_0 = const()[name = tensor<string, []>("op_907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
599 tensor<fp16, [1, 64, 1, 375]> var_907_cast_fp16 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = var_737_cast_fp16)[name = tensor<string, []>("op_907_cast_fp16")];
600 tensor<int32, [4]> k_3_perm_0 = const()[name = tensor<string, []>("k_3_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
601 tensor<int32, [4]> var_912_begin_0 = const()[name = tensor<string, []>("op_912_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
602 tensor<int32, [4]> var_912_end_0 = const()[name = tensor<string, []>("op_912_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 64])];
603 tensor<bool, [4]> var_912_end_mask_0 = const()[name = tensor<string, []>("op_912_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
604 tensor<fp16, [1, 1500, 1, 384]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = key_3_cast_fp16)[name = tensor<string, []>("transpose_2")];
605 tensor<fp16, [1, 1500, 1, 64]> var_912_cast_fp16 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_912_cast_fp16")];
606 tensor<int32, [4]> var_916_begin_0 = const()[name = tensor<string, []>("op_916_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
607 tensor<int32, [4]> var_916_end_0 = const()[name = tensor<string, []>("op_916_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 128])];
608 tensor<bool, [4]> var_916_end_mask_0 = const()[name = tensor<string, []>("op_916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
609 tensor<fp16, [1, 1500, 1, 64]> var_916_cast_fp16 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_916_cast_fp16")];
610 tensor<int32, [4]> var_920_begin_0 = const()[name = tensor<string, []>("op_920_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
611 tensor<int32, [4]> var_920_end_0 = const()[name = tensor<string, []>("op_920_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 192])];
612 tensor<bool, [4]> var_920_end_mask_0 = const()[name = tensor<string, []>("op_920_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
613 tensor<fp16, [1, 1500, 1, 64]> var_920_cast_fp16 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_920_cast_fp16")];
614 tensor<int32, [4]> var_924_begin_0 = const()[name = tensor<string, []>("op_924_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 192])];
615 tensor<int32, [4]> var_924_end_0 = const()[name = tensor<string, []>("op_924_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 256])];
616 tensor<bool, [4]> var_924_end_mask_0 = const()[name = tensor<string, []>("op_924_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
617 tensor<fp16, [1, 1500, 1, 64]> var_924_cast_fp16 = slice_by_index(begin = var_924_begin_0, end = var_924_end_0, end_mask = var_924_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_924_cast_fp16")];
618 tensor<int32, [4]> var_928_begin_0 = const()[name = tensor<string, []>("op_928_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
619 tensor<int32, [4]> var_928_end_0 = const()[name = tensor<string, []>("op_928_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 320])];
620 tensor<bool, [4]> var_928_end_mask_0 = const()[name = tensor<string, []>("op_928_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
621 tensor<fp16, [1, 1500, 1, 64]> var_928_cast_fp16 = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_928_cast_fp16")];
622 tensor<int32, [4]> var_932_begin_0 = const()[name = tensor<string, []>("op_932_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 320])];
623 tensor<int32, [4]> var_932_end_0 = const()[name = tensor<string, []>("op_932_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 384])];
624 tensor<bool, [4]> var_932_end_mask_0 = const()[name = tensor<string, []>("op_932_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
625 tensor<fp16, [1, 1500, 1, 64]> var_932_cast_fp16 = slice_by_index(begin = var_932_begin_0, end = var_932_end_0, end_mask = var_932_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_932_cast_fp16")];
626 tensor<int32, [4]> var_934_begin_0 = const()[name = tensor<string, []>("op_934_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
627 tensor<int32, [4]> var_934_end_0 = const()[name = tensor<string, []>("op_934_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
628 tensor<bool, [4]> var_934_end_mask_0 = const()[name = tensor<string, []>("op_934_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
629 tensor<fp16, [1, 64, 1, 1500]> var_934_cast_fp16 = slice_by_index(begin = var_934_begin_0, end = var_934_end_0, end_mask = var_934_end_mask_0, x = value_3_cast_fp16)[name = tensor<string, []>("op_934_cast_fp16")];
630 tensor<int32, [4]> var_938_begin_0 = const()[name = tensor<string, []>("op_938_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
631 tensor<int32, [4]> var_938_end_0 = const()[name = tensor<string, []>("op_938_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
632 tensor<bool, [4]> var_938_end_mask_0 = const()[name = tensor<string, []>("op_938_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
633 tensor<fp16, [1, 64, 1, 1500]> var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = value_3_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
634 tensor<int32, [4]> var_942_begin_0 = const()[name = tensor<string, []>("op_942_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
635 tensor<int32, [4]> var_942_end_0 = const()[name = tensor<string, []>("op_942_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
636 tensor<bool, [4]> var_942_end_mask_0 = const()[name = tensor<string, []>("op_942_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
637 tensor<fp16, [1, 64, 1, 1500]> var_942_cast_fp16 = slice_by_index(begin = var_942_begin_0, end = var_942_end_0, end_mask = var_942_end_mask_0, x = value_3_cast_fp16)[name = tensor<string, []>("op_942_cast_fp16")];
638 tensor<int32, [4]> var_946_begin_0 = const()[name = tensor<string, []>("op_946_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
639 tensor<int32, [4]> var_946_end_0 = const()[name = tensor<string, []>("op_946_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
640 tensor<bool, [4]> var_946_end_mask_0 = const()[name = tensor<string, []>("op_946_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
641 tensor<fp16, [1, 64, 1, 1500]> var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = value_3_cast_fp16)[name = tensor<string, []>("op_946_cast_fp16")];
642 tensor<int32, [4]> var_950_begin_0 = const()[name = tensor<string, []>("op_950_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
643 tensor<int32, [4]> var_950_end_0 = const()[name = tensor<string, []>("op_950_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
644 tensor<bool, [4]> var_950_end_mask_0 = const()[name = tensor<string, []>("op_950_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
645 tensor<fp16, [1, 64, 1, 1500]> var_950_cast_fp16 = slice_by_index(begin = var_950_begin_0, end = var_950_end_0, end_mask = var_950_end_mask_0, x = value_3_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
646 tensor<int32, [4]> var_954_begin_0 = const()[name = tensor<string, []>("op_954_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
647 tensor<int32, [4]> var_954_end_0 = const()[name = tensor<string, []>("op_954_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
648 tensor<bool, [4]> var_954_end_mask_0 = const()[name = tensor<string, []>("op_954_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
649 tensor<fp16, [1, 64, 1, 1500]> var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = value_3_cast_fp16)[name = tensor<string, []>("op_954_cast_fp16")];
650 tensor<string, []> _SplitHeadsQ__mh_w_49_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_49_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
651 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_49_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_49_equation_0, values = (var_912_cast_fp16, var_746_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_49_cast_fp16")];
652 tensor<string, []> _SplitHeadsQ__mh_w_51_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_51_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
653 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_51_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_51_equation_0, values = (var_912_cast_fp16, var_753_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_51_cast_fp16")];
654 tensor<string, []> _SplitHeadsQ__mh_w_53_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_53_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
655 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_53_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_53_equation_0, values = (var_912_cast_fp16, var_760_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_53_cast_fp16")];
656 tensor<string, []> _SplitHeadsQ__mh_w_55_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_55_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
657 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_55_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_55_equation_0, values = (var_912_cast_fp16, var_767_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_55_cast_fp16")];
658 tensor<string, []> _SplitHeadsQ__mh_w_57_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_57_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
659 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_57_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_57_equation_0, values = (var_916_cast_fp16, var_774_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_57_cast_fp16")];
660 tensor<string, []> _SplitHeadsQ__mh_w_59_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_59_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
661 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_59_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_59_equation_0, values = (var_916_cast_fp16, var_781_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_59_cast_fp16")];
662 tensor<string, []> _SplitHeadsQ__mh_w_61_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_61_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
663 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_61_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_61_equation_0, values = (var_916_cast_fp16, var_788_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_61_cast_fp16")];
664 tensor<string, []> _SplitHeadsQ__mh_w_63_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_63_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
665 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_63_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_63_equation_0, values = (var_916_cast_fp16, var_795_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_63_cast_fp16")];
666 tensor<string, []> _SplitHeadsQ__mh_w_65_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_65_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
667 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_65_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_65_equation_0, values = (var_920_cast_fp16, var_802_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_65_cast_fp16")];
668 tensor<string, []> _SplitHeadsQ__mh_w_67_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_67_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
669 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_67_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_67_equation_0, values = (var_920_cast_fp16, var_809_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_67_cast_fp16")];
670 tensor<string, []> _SplitHeadsQ__mh_w_69_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_69_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
671 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_69_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_69_equation_0, values = (var_920_cast_fp16, var_816_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_69_cast_fp16")];
672 tensor<string, []> _SplitHeadsQ__mh_w_71_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_71_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
673 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_71_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_71_equation_0, values = (var_920_cast_fp16, var_823_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_71_cast_fp16")];
674 tensor<string, []> _SplitHeadsQ__mh_w_73_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_73_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
675 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_73_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_73_equation_0, values = (var_924_cast_fp16, var_830_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_73_cast_fp16")];
676 tensor<string, []> _SplitHeadsQ__mh_w_75_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_75_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
677 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_75_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_75_equation_0, values = (var_924_cast_fp16, var_837_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_75_cast_fp16")];
678 tensor<string, []> _SplitHeadsQ__mh_w_77_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_77_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
679 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_77_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_77_equation_0, values = (var_924_cast_fp16, var_844_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_77_cast_fp16")];
680 tensor<string, []> _SplitHeadsQ__mh_w_79_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_79_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
681 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_79_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_79_equation_0, values = (var_924_cast_fp16, var_851_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_79_cast_fp16")];
682 tensor<string, []> _SplitHeadsQ__mh_w_81_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_81_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
683 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_81_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_81_equation_0, values = (var_928_cast_fp16, var_858_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_81_cast_fp16")];
684 tensor<string, []> _SplitHeadsQ__mh_w_83_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_83_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
685 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_83_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_83_equation_0, values = (var_928_cast_fp16, var_865_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_83_cast_fp16")];
686 tensor<string, []> _SplitHeadsQ__mh_w_85_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_85_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
687 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_85_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_85_equation_0, values = (var_928_cast_fp16, var_872_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_85_cast_fp16")];
688 tensor<string, []> _SplitHeadsQ__mh_w_87_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_87_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
689 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_87_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_87_equation_0, values = (var_928_cast_fp16, var_879_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_87_cast_fp16")];
690 tensor<string, []> _SplitHeadsQ__mh_w_89_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_89_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
691 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_89_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_89_equation_0, values = (var_932_cast_fp16, var_886_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_89_cast_fp16")];
692 tensor<string, []> _SplitHeadsQ__mh_w_91_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_91_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
693 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_91_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_91_equation_0, values = (var_932_cast_fp16, var_893_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_91_cast_fp16")];
694 tensor<string, []> _SplitHeadsQ__mh_w_93_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_93_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
695 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_93_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_93_equation_0, values = (var_932_cast_fp16, var_900_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_93_cast_fp16")];
696 tensor<string, []> _SplitHeadsQ__mh_w_95_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_95_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
697 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_95_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_95_equation_0, values = (var_932_cast_fp16, var_907_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_95_cast_fp16")];
698 tensor<fp16, []> var_1005_to_fp16 = const()[name = tensor<string, []>("op_1005_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
699 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_49_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_49_cast_fp16, y = var_1005_to_fp16)[name = tensor<string, []>("aw_chunk_49_cast_fp16")];
700 tensor<fp16, []> var_1007_to_fp16 = const()[name = tensor<string, []>("op_1007_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
701 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_51_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_51_cast_fp16, y = var_1007_to_fp16)[name = tensor<string, []>("aw_chunk_51_cast_fp16")];
702 tensor<fp16, []> var_1009_to_fp16 = const()[name = tensor<string, []>("op_1009_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
703 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_53_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_53_cast_fp16, y = var_1009_to_fp16)[name = tensor<string, []>("aw_chunk_53_cast_fp16")];
704 tensor<fp16, []> var_1011_to_fp16 = const()[name = tensor<string, []>("op_1011_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
705 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_55_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_55_cast_fp16, y = var_1011_to_fp16)[name = tensor<string, []>("aw_chunk_55_cast_fp16")];
706 tensor<fp16, []> var_1013_to_fp16 = const()[name = tensor<string, []>("op_1013_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
707 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_57_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_57_cast_fp16, y = var_1013_to_fp16)[name = tensor<string, []>("aw_chunk_57_cast_fp16")];
708 tensor<fp16, []> var_1015_to_fp16 = const()[name = tensor<string, []>("op_1015_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
709 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_59_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_59_cast_fp16, y = var_1015_to_fp16)[name = tensor<string, []>("aw_chunk_59_cast_fp16")];
710 tensor<fp16, []> var_1017_to_fp16 = const()[name = tensor<string, []>("op_1017_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
711 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_61_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_61_cast_fp16, y = var_1017_to_fp16)[name = tensor<string, []>("aw_chunk_61_cast_fp16")];
712 tensor<fp16, []> var_1019_to_fp16 = const()[name = tensor<string, []>("op_1019_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
713 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_63_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_63_cast_fp16, y = var_1019_to_fp16)[name = tensor<string, []>("aw_chunk_63_cast_fp16")];
714 tensor<fp16, []> var_1021_to_fp16 = const()[name = tensor<string, []>("op_1021_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
715 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_65_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_65_cast_fp16, y = var_1021_to_fp16)[name = tensor<string, []>("aw_chunk_65_cast_fp16")];
716 tensor<fp16, []> var_1023_to_fp16 = const()[name = tensor<string, []>("op_1023_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
717 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_67_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_67_cast_fp16, y = var_1023_to_fp16)[name = tensor<string, []>("aw_chunk_67_cast_fp16")];
718 tensor<fp16, []> var_1025_to_fp16 = const()[name = tensor<string, []>("op_1025_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
719 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_69_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_69_cast_fp16, y = var_1025_to_fp16)[name = tensor<string, []>("aw_chunk_69_cast_fp16")];
720 tensor<fp16, []> var_1027_to_fp16 = const()[name = tensor<string, []>("op_1027_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
721 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_71_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_71_cast_fp16, y = var_1027_to_fp16)[name = tensor<string, []>("aw_chunk_71_cast_fp16")];
722 tensor<fp16, []> var_1029_to_fp16 = const()[name = tensor<string, []>("op_1029_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
723 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_73_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_73_cast_fp16, y = var_1029_to_fp16)[name = tensor<string, []>("aw_chunk_73_cast_fp16")];
724 tensor<fp16, []> var_1031_to_fp16 = const()[name = tensor<string, []>("op_1031_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
725 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_75_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_75_cast_fp16, y = var_1031_to_fp16)[name = tensor<string, []>("aw_chunk_75_cast_fp16")];
726 tensor<fp16, []> var_1033_to_fp16 = const()[name = tensor<string, []>("op_1033_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
727 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_77_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_77_cast_fp16, y = var_1033_to_fp16)[name = tensor<string, []>("aw_chunk_77_cast_fp16")];
728 tensor<fp16, []> var_1035_to_fp16 = const()[name = tensor<string, []>("op_1035_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
729 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_79_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_79_cast_fp16, y = var_1035_to_fp16)[name = tensor<string, []>("aw_chunk_79_cast_fp16")];
730 tensor<fp16, []> var_1037_to_fp16 = const()[name = tensor<string, []>("op_1037_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
731 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_81_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_81_cast_fp16, y = var_1037_to_fp16)[name = tensor<string, []>("aw_chunk_81_cast_fp16")];
732 tensor<fp16, []> var_1039_to_fp16 = const()[name = tensor<string, []>("op_1039_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
733 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_83_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_83_cast_fp16, y = var_1039_to_fp16)[name = tensor<string, []>("aw_chunk_83_cast_fp16")];
734 tensor<fp16, []> var_1041_to_fp16 = const()[name = tensor<string, []>("op_1041_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
735 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_85_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_85_cast_fp16, y = var_1041_to_fp16)[name = tensor<string, []>("aw_chunk_85_cast_fp16")];
736 tensor<fp16, []> var_1043_to_fp16 = const()[name = tensor<string, []>("op_1043_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
737 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_87_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_87_cast_fp16, y = var_1043_to_fp16)[name = tensor<string, []>("aw_chunk_87_cast_fp16")];
738 tensor<fp16, []> var_1045_to_fp16 = const()[name = tensor<string, []>("op_1045_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
739 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_89_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_89_cast_fp16, y = var_1045_to_fp16)[name = tensor<string, []>("aw_chunk_89_cast_fp16")];
740 tensor<fp16, []> var_1047_to_fp16 = const()[name = tensor<string, []>("op_1047_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
741 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_91_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_91_cast_fp16, y = var_1047_to_fp16)[name = tensor<string, []>("aw_chunk_91_cast_fp16")];
742 tensor<fp16, []> var_1049_to_fp16 = const()[name = tensor<string, []>("op_1049_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
743 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_93_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_93_cast_fp16, y = var_1049_to_fp16)[name = tensor<string, []>("aw_chunk_93_cast_fp16")];
744 tensor<fp16, []> var_1051_to_fp16 = const()[name = tensor<string, []>("op_1051_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
745 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_95_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_95_cast_fp16, y = var_1051_to_fp16)[name = tensor<string, []>("aw_chunk_95_cast_fp16")];
746 tensor<fp16, [1, 1500, 1, 375]> var_1053_cast_fp16 = softmax(axis = var_662, x = aw_chunk_49_cast_fp16)[name = tensor<string, []>("op_1053_cast_fp16")];
747 tensor<fp16, [1, 1500, 1, 375]> var_1054_cast_fp16 = softmax(axis = var_662, x = aw_chunk_51_cast_fp16)[name = tensor<string, []>("op_1054_cast_fp16")];
748 tensor<fp16, [1, 1500, 1, 375]> var_1055_cast_fp16 = softmax(axis = var_662, x = aw_chunk_53_cast_fp16)[name = tensor<string, []>("op_1055_cast_fp16")];
749 tensor<fp16, [1, 1500, 1, 375]> var_1056_cast_fp16 = softmax(axis = var_662, x = aw_chunk_55_cast_fp16)[name = tensor<string, []>("op_1056_cast_fp16")];
750 tensor<fp16, [1, 1500, 1, 375]> var_1057_cast_fp16 = softmax(axis = var_662, x = aw_chunk_57_cast_fp16)[name = tensor<string, []>("op_1057_cast_fp16")];
751 tensor<fp16, [1, 1500, 1, 375]> var_1058_cast_fp16 = softmax(axis = var_662, x = aw_chunk_59_cast_fp16)[name = tensor<string, []>("op_1058_cast_fp16")];
752 tensor<fp16, [1, 1500, 1, 375]> var_1059_cast_fp16 = softmax(axis = var_662, x = aw_chunk_61_cast_fp16)[name = tensor<string, []>("op_1059_cast_fp16")];
753 tensor<fp16, [1, 1500, 1, 375]> var_1060_cast_fp16 = softmax(axis = var_662, x = aw_chunk_63_cast_fp16)[name = tensor<string, []>("op_1060_cast_fp16")];
754 tensor<fp16, [1, 1500, 1, 375]> var_1061_cast_fp16 = softmax(axis = var_662, x = aw_chunk_65_cast_fp16)[name = tensor<string, []>("op_1061_cast_fp16")];
755 tensor<fp16, [1, 1500, 1, 375]> var_1062_cast_fp16 = softmax(axis = var_662, x = aw_chunk_67_cast_fp16)[name = tensor<string, []>("op_1062_cast_fp16")];
756 tensor<fp16, [1, 1500, 1, 375]> var_1063_cast_fp16 = softmax(axis = var_662, x = aw_chunk_69_cast_fp16)[name = tensor<string, []>("op_1063_cast_fp16")];
757 tensor<fp16, [1, 1500, 1, 375]> var_1064_cast_fp16 = softmax(axis = var_662, x = aw_chunk_71_cast_fp16)[name = tensor<string, []>("op_1064_cast_fp16")];
758 tensor<fp16, [1, 1500, 1, 375]> var_1065_cast_fp16 = softmax(axis = var_662, x = aw_chunk_73_cast_fp16)[name = tensor<string, []>("op_1065_cast_fp16")];
759 tensor<fp16, [1, 1500, 1, 375]> var_1066_cast_fp16 = softmax(axis = var_662, x = aw_chunk_75_cast_fp16)[name = tensor<string, []>("op_1066_cast_fp16")];
760 tensor<fp16, [1, 1500, 1, 375]> var_1067_cast_fp16 = softmax(axis = var_662, x = aw_chunk_77_cast_fp16)[name = tensor<string, []>("op_1067_cast_fp16")];
761 tensor<fp16, [1, 1500, 1, 375]> var_1068_cast_fp16 = softmax(axis = var_662, x = aw_chunk_79_cast_fp16)[name = tensor<string, []>("op_1068_cast_fp16")];
762 tensor<fp16, [1, 1500, 1, 375]> var_1069_cast_fp16 = softmax(axis = var_662, x = aw_chunk_81_cast_fp16)[name = tensor<string, []>("op_1069_cast_fp16")];
763 tensor<fp16, [1, 1500, 1, 375]> var_1070_cast_fp16 = softmax(axis = var_662, x = aw_chunk_83_cast_fp16)[name = tensor<string, []>("op_1070_cast_fp16")];
764 tensor<fp16, [1, 1500, 1, 375]> var_1071_cast_fp16 = softmax(axis = var_662, x = aw_chunk_85_cast_fp16)[name = tensor<string, []>("op_1071_cast_fp16")];
765 tensor<fp16, [1, 1500, 1, 375]> var_1072_cast_fp16 = softmax(axis = var_662, x = aw_chunk_87_cast_fp16)[name = tensor<string, []>("op_1072_cast_fp16")];
766 tensor<fp16, [1, 1500, 1, 375]> var_1073_cast_fp16 = softmax(axis = var_662, x = aw_chunk_89_cast_fp16)[name = tensor<string, []>("op_1073_cast_fp16")];
767 tensor<fp16, [1, 1500, 1, 375]> var_1074_cast_fp16 = softmax(axis = var_662, x = aw_chunk_91_cast_fp16)[name = tensor<string, []>("op_1074_cast_fp16")];
768 tensor<fp16, [1, 1500, 1, 375]> var_1075_cast_fp16 = softmax(axis = var_662, x = aw_chunk_93_cast_fp16)[name = tensor<string, []>("op_1075_cast_fp16")];
769 tensor<fp16, [1, 1500, 1, 375]> var_1076_cast_fp16 = softmax(axis = var_662, x = aw_chunk_95_cast_fp16)[name = tensor<string, []>("op_1076_cast_fp16")];
770 tensor<string, []> var_1078_equation_0 = const()[name = tensor<string, []>("op_1078_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
771 tensor<fp16, [1, 64, 1, 375]> var_1078_cast_fp16 = einsum(equation = var_1078_equation_0, values = (var_934_cast_fp16, var_1053_cast_fp16))[name = tensor<string, []>("op_1078_cast_fp16")];
772 tensor<string, []> var_1080_equation_0 = const()[name = tensor<string, []>("op_1080_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
773 tensor<fp16, [1, 64, 1, 375]> var_1080_cast_fp16 = einsum(equation = var_1080_equation_0, values = (var_934_cast_fp16, var_1054_cast_fp16))[name = tensor<string, []>("op_1080_cast_fp16")];
774 tensor<string, []> var_1082_equation_0 = const()[name = tensor<string, []>("op_1082_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
775 tensor<fp16, [1, 64, 1, 375]> var_1082_cast_fp16 = einsum(equation = var_1082_equation_0, values = (var_934_cast_fp16, var_1055_cast_fp16))[name = tensor<string, []>("op_1082_cast_fp16")];
776 tensor<string, []> var_1084_equation_0 = const()[name = tensor<string, []>("op_1084_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
777 tensor<fp16, [1, 64, 1, 375]> var_1084_cast_fp16 = einsum(equation = var_1084_equation_0, values = (var_934_cast_fp16, var_1056_cast_fp16))[name = tensor<string, []>("op_1084_cast_fp16")];
778 tensor<string, []> var_1086_equation_0 = const()[name = tensor<string, []>("op_1086_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
779 tensor<fp16, [1, 64, 1, 375]> var_1086_cast_fp16 = einsum(equation = var_1086_equation_0, values = (var_938_cast_fp16, var_1057_cast_fp16))[name = tensor<string, []>("op_1086_cast_fp16")];
780 tensor<string, []> var_1088_equation_0 = const()[name = tensor<string, []>("op_1088_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
781 tensor<fp16, [1, 64, 1, 375]> var_1088_cast_fp16 = einsum(equation = var_1088_equation_0, values = (var_938_cast_fp16, var_1058_cast_fp16))[name = tensor<string, []>("op_1088_cast_fp16")];
782 tensor<string, []> var_1090_equation_0 = const()[name = tensor<string, []>("op_1090_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
783 tensor<fp16, [1, 64, 1, 375]> var_1090_cast_fp16 = einsum(equation = var_1090_equation_0, values = (var_938_cast_fp16, var_1059_cast_fp16))[name = tensor<string, []>("op_1090_cast_fp16")];
784 tensor<string, []> var_1092_equation_0 = const()[name = tensor<string, []>("op_1092_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
785 tensor<fp16, [1, 64, 1, 375]> var_1092_cast_fp16 = einsum(equation = var_1092_equation_0, values = (var_938_cast_fp16, var_1060_cast_fp16))[name = tensor<string, []>("op_1092_cast_fp16")];
786 tensor<string, []> var_1094_equation_0 = const()[name = tensor<string, []>("op_1094_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
787 tensor<fp16, [1, 64, 1, 375]> var_1094_cast_fp16 = einsum(equation = var_1094_equation_0, values = (var_942_cast_fp16, var_1061_cast_fp16))[name = tensor<string, []>("op_1094_cast_fp16")];
788 tensor<string, []> var_1096_equation_0 = const()[name = tensor<string, []>("op_1096_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
789 tensor<fp16, [1, 64, 1, 375]> var_1096_cast_fp16 = einsum(equation = var_1096_equation_0, values = (var_942_cast_fp16, var_1062_cast_fp16))[name = tensor<string, []>("op_1096_cast_fp16")];
790 tensor<string, []> var_1098_equation_0 = const()[name = tensor<string, []>("op_1098_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
791 tensor<fp16, [1, 64, 1, 375]> var_1098_cast_fp16 = einsum(equation = var_1098_equation_0, values = (var_942_cast_fp16, var_1063_cast_fp16))[name = tensor<string, []>("op_1098_cast_fp16")];
792 tensor<string, []> var_1100_equation_0 = const()[name = tensor<string, []>("op_1100_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
793 tensor<fp16, [1, 64, 1, 375]> var_1100_cast_fp16 = einsum(equation = var_1100_equation_0, values = (var_942_cast_fp16, var_1064_cast_fp16))[name = tensor<string, []>("op_1100_cast_fp16")];
794 tensor<string, []> var_1102_equation_0 = const()[name = tensor<string, []>("op_1102_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
795 tensor<fp16, [1, 64, 1, 375]> var_1102_cast_fp16 = einsum(equation = var_1102_equation_0, values = (var_946_cast_fp16, var_1065_cast_fp16))[name = tensor<string, []>("op_1102_cast_fp16")];
796 tensor<string, []> var_1104_equation_0 = const()[name = tensor<string, []>("op_1104_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
797 tensor<fp16, [1, 64, 1, 375]> var_1104_cast_fp16 = einsum(equation = var_1104_equation_0, values = (var_946_cast_fp16, var_1066_cast_fp16))[name = tensor<string, []>("op_1104_cast_fp16")];
798 tensor<string, []> var_1106_equation_0 = const()[name = tensor<string, []>("op_1106_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
799 tensor<fp16, [1, 64, 1, 375]> var_1106_cast_fp16 = einsum(equation = var_1106_equation_0, values = (var_946_cast_fp16, var_1067_cast_fp16))[name = tensor<string, []>("op_1106_cast_fp16")];
800 tensor<string, []> var_1108_equation_0 = const()[name = tensor<string, []>("op_1108_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
801 tensor<fp16, [1, 64, 1, 375]> var_1108_cast_fp16 = einsum(equation = var_1108_equation_0, values = (var_946_cast_fp16, var_1068_cast_fp16))[name = tensor<string, []>("op_1108_cast_fp16")];
802 tensor<string, []> var_1110_equation_0 = const()[name = tensor<string, []>("op_1110_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
803 tensor<fp16, [1, 64, 1, 375]> var_1110_cast_fp16 = einsum(equation = var_1110_equation_0, values = (var_950_cast_fp16, var_1069_cast_fp16))[name = tensor<string, []>("op_1110_cast_fp16")];
804 tensor<string, []> var_1112_equation_0 = const()[name = tensor<string, []>("op_1112_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
805 tensor<fp16, [1, 64, 1, 375]> var_1112_cast_fp16 = einsum(equation = var_1112_equation_0, values = (var_950_cast_fp16, var_1070_cast_fp16))[name = tensor<string, []>("op_1112_cast_fp16")];
806 tensor<string, []> var_1114_equation_0 = const()[name = tensor<string, []>("op_1114_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
807 tensor<fp16, [1, 64, 1, 375]> var_1114_cast_fp16 = einsum(equation = var_1114_equation_0, values = (var_950_cast_fp16, var_1071_cast_fp16))[name = tensor<string, []>("op_1114_cast_fp16")];
808 tensor<string, []> var_1116_equation_0 = const()[name = tensor<string, []>("op_1116_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
809 tensor<fp16, [1, 64, 1, 375]> var_1116_cast_fp16 = einsum(equation = var_1116_equation_0, values = (var_950_cast_fp16, var_1072_cast_fp16))[name = tensor<string, []>("op_1116_cast_fp16")];
810 tensor<string, []> var_1118_equation_0 = const()[name = tensor<string, []>("op_1118_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
811 tensor<fp16, [1, 64, 1, 375]> var_1118_cast_fp16 = einsum(equation = var_1118_equation_0, values = (var_954_cast_fp16, var_1073_cast_fp16))[name = tensor<string, []>("op_1118_cast_fp16")];
812 tensor<string, []> var_1120_equation_0 = const()[name = tensor<string, []>("op_1120_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
813 tensor<fp16, [1, 64, 1, 375]> var_1120_cast_fp16 = einsum(equation = var_1120_equation_0, values = (var_954_cast_fp16, var_1074_cast_fp16))[name = tensor<string, []>("op_1120_cast_fp16")];
814 tensor<string, []> var_1122_equation_0 = const()[name = tensor<string, []>("op_1122_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
815 tensor<fp16, [1, 64, 1, 375]> var_1122_cast_fp16 = einsum(equation = var_1122_equation_0, values = (var_954_cast_fp16, var_1075_cast_fp16))[name = tensor<string, []>("op_1122_cast_fp16")];
816 tensor<string, []> var_1124_equation_0 = const()[name = tensor<string, []>("op_1124_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
817 tensor<fp16, [1, 64, 1, 375]> var_1124_cast_fp16 = einsum(equation = var_1124_equation_0, values = (var_954_cast_fp16, var_1076_cast_fp16))[name = tensor<string, []>("op_1124_cast_fp16")];
818 tensor<bool, []> var_1126_interleave_0 = const()[name = tensor<string, []>("op_1126_interleave_0"), val = tensor<bool, []>(false)];
819 tensor<fp16, [1, 64, 1, 1500]> var_1126_cast_fp16 = concat(axis = var_651, interleave = var_1126_interleave_0, values = (var_1078_cast_fp16, var_1080_cast_fp16, var_1082_cast_fp16, var_1084_cast_fp16))[name = tensor<string, []>("op_1126_cast_fp16")];
820 tensor<bool, []> var_1128_interleave_0 = const()[name = tensor<string, []>("op_1128_interleave_0"), val = tensor<bool, []>(false)];
821 tensor<fp16, [1, 64, 1, 1500]> var_1128_cast_fp16 = concat(axis = var_651, interleave = var_1128_interleave_0, values = (var_1086_cast_fp16, var_1088_cast_fp16, var_1090_cast_fp16, var_1092_cast_fp16))[name = tensor<string, []>("op_1128_cast_fp16")];
822 tensor<bool, []> var_1130_interleave_0 = const()[name = tensor<string, []>("op_1130_interleave_0"), val = tensor<bool, []>(false)];
823 tensor<fp16, [1, 64, 1, 1500]> var_1130_cast_fp16 = concat(axis = var_651, interleave = var_1130_interleave_0, values = (var_1094_cast_fp16, var_1096_cast_fp16, var_1098_cast_fp16, var_1100_cast_fp16))[name = tensor<string, []>("op_1130_cast_fp16")];
824 tensor<bool, []> var_1132_interleave_0 = const()[name = tensor<string, []>("op_1132_interleave_0"), val = tensor<bool, []>(false)];
825 tensor<fp16, [1, 64, 1, 1500]> var_1132_cast_fp16 = concat(axis = var_651, interleave = var_1132_interleave_0, values = (var_1102_cast_fp16, var_1104_cast_fp16, var_1106_cast_fp16, var_1108_cast_fp16))[name = tensor<string, []>("op_1132_cast_fp16")];
826 tensor<bool, []> var_1134_interleave_0 = const()[name = tensor<string, []>("op_1134_interleave_0"), val = tensor<bool, []>(false)];
827 tensor<fp16, [1, 64, 1, 1500]> var_1134_cast_fp16 = concat(axis = var_651, interleave = var_1134_interleave_0, values = (var_1110_cast_fp16, var_1112_cast_fp16, var_1114_cast_fp16, var_1116_cast_fp16))[name = tensor<string, []>("op_1134_cast_fp16")];
828 tensor<bool, []> var_1136_interleave_0 = const()[name = tensor<string, []>("op_1136_interleave_0"), val = tensor<bool, []>(false)];
829 tensor<fp16, [1, 64, 1, 1500]> var_1136_cast_fp16 = concat(axis = var_651, interleave = var_1136_interleave_0, values = (var_1118_cast_fp16, var_1120_cast_fp16, var_1122_cast_fp16, var_1124_cast_fp16))[name = tensor<string, []>("op_1136_cast_fp16")];
830 tensor<bool, []> input_9_interleave_0 = const()[name = tensor<string, []>("input_9_interleave_0"), val = tensor<bool, []>(false)];
831 tensor<fp16, [1, 384, 1, 1500]> input_9_cast_fp16 = concat(axis = var_662, interleave = input_9_interleave_0, values = (var_1126_cast_fp16, var_1128_cast_fp16, var_1130_cast_fp16, var_1132_cast_fp16, var_1134_cast_fp16, var_1136_cast_fp16))[name = tensor<string, []>("input_9_cast_fp16")];
832 tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("valid")];
833 tensor<int32, [2]> obj_7_strides_0 = const()[name = tensor<string, []>("obj_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
834 tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
835 tensor<int32, [2]> obj_7_dilations_0 = const()[name = tensor<string, []>("obj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
836 tensor<int32, []> obj_7_groups_0 = const()[name = tensor<string, []>("obj_7_groups_0"), val = tensor<int32, []>(1)];
837 tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6662016)))];
838 tensor<fp16, [384]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6956992)))];
839 tensor<fp16, [1, 384, 1, 1500]> obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
840 tensor<fp16, [1, 384, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
841 tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
842 tensor<fp16, []> var_1155_to_fp16 = const()[name = tensor<string, []>("op_1155_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
843 tensor<fp16, [1, 384, 1, 1500]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_1155_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
844 tensor<fp16, [384]> input_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_11_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6957824)))];
845 tensor<fp16, [384]> input_11_beta_0_to_fp16 = const()[name = tensor<string, []>("input_11_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6958656)))];
846 tensor<fp16, []> input_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
847 tensor<fp16, [1, 384, 1, 1500]> input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_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, []>("input_11_cast_fp16")];
848 tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
849 tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
850 tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
851 tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
852 tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
853 tensor<fp16, [1536, 384, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6959488)))];
854 tensor<fp16, [1536]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8139200)))];
855 tensor<fp16, [1, 1536, 1, 1500]> input_13_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, 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_1_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
856 tensor<string, []> input_15_mode_0 = const()[name = tensor<string, []>("input_15_mode_0"), val = tensor<string, []>("EXACT")];
857 tensor<fp16, [1, 1536, 1, 1500]> input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
858 tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("valid")];
859 tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = tensor<string, []>("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
860 tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
861 tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = tensor<string, []>("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
862 tensor<int32, []> hidden_states_7_groups_0 = const()[name = tensor<string, []>("hidden_states_7_groups_0"), val = tensor<int32, []>(1)];
863 tensor<fp16, [384, 1536, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8142336)))];
864 tensor<fp16, [384]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322048)))];
865 tensor<fp16, [1, 384, 1, 1500]> hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
866 tensor<fp16, [1, 384, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
867 tensor<int32, []> var_1184 = const()[name = tensor<string, []>("op_1184"), val = tensor<int32, []>(3)];
868 tensor<int32, []> var_1195 = const()[name = tensor<string, []>("op_1195"), val = tensor<int32, []>(1)];
869 tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
870 tensor<fp16, []> var_1212_to_fp16 = const()[name = tensor<string, []>("op_1212_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
871 tensor<fp16, [1, 384, 1, 1500]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1212_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
872 tensor<fp16, [384]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322880)))];
873 tensor<fp16, [384]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9323712)))];
874 tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
875 tensor<fp16, [1, 384, 1, 1500]> 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_9_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
876 tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")];
877 tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
878 tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
879 tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
880 tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)];
881 tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9324544)))];
882 tensor<fp16, [384]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9619520)))];
883 tensor<fp16, [1, 384, 1, 1500]> query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
884 tensor<string, []> key_5_pad_type_0 = const()[name = tensor<string, []>("key_5_pad_type_0"), val = tensor<string, []>("valid")];
885 tensor<int32, [2]> key_5_strides_0 = const()[name = tensor<string, []>("key_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
886 tensor<int32, [4]> key_5_pad_0 = const()[name = tensor<string, []>("key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
887 tensor<int32, [2]> key_5_dilations_0 = const()[name = tensor<string, []>("key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
888 tensor<int32, []> key_5_groups_0 = const()[name = tensor<string, []>("key_5_groups_0"), val = tensor<int32, []>(1)];
889 tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9620352)))];
890 tensor<fp16, [1, 384, 1, 1500]> key_5_cast_fp16 = conv(dilations = key_5_dilations_0, groups = key_5_groups_0, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
891 tensor<string, []> value_5_pad_type_0 = const()[name = tensor<string, []>("value_5_pad_type_0"), val = tensor<string, []>("valid")];
892 tensor<int32, [2]> value_5_strides_0 = const()[name = tensor<string, []>("value_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
893 tensor<int32, [4]> value_5_pad_0 = const()[name = tensor<string, []>("value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
894 tensor<int32, [2]> value_5_dilations_0 = const()[name = tensor<string, []>("value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
895 tensor<int32, []> value_5_groups_0 = const()[name = tensor<string, []>("value_5_groups_0"), val = tensor<int32, []>(1)];
896 tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9915328)))];
897 tensor<fp16, [384]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10210304)))];
898 tensor<fp16, [1, 384, 1, 1500]> value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
899 tensor<int32, [4]> var_1250_begin_0 = const()[name = tensor<string, []>("op_1250_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
900 tensor<int32, [4]> var_1250_end_0 = const()[name = tensor<string, []>("op_1250_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
901 tensor<bool, [4]> var_1250_end_mask_0 = const()[name = tensor<string, []>("op_1250_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
902 tensor<fp16, [1, 64, 1, 1500]> var_1250_cast_fp16 = slice_by_index(begin = var_1250_begin_0, end = var_1250_end_0, end_mask = var_1250_end_mask_0, x = query_5_cast_fp16)[name = tensor<string, []>("op_1250_cast_fp16")];
903 tensor<int32, [4]> var_1254_begin_0 = const()[name = tensor<string, []>("op_1254_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
904 tensor<int32, [4]> var_1254_end_0 = const()[name = tensor<string, []>("op_1254_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
905 tensor<bool, [4]> var_1254_end_mask_0 = const()[name = tensor<string, []>("op_1254_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
906 tensor<fp16, [1, 64, 1, 1500]> var_1254_cast_fp16 = slice_by_index(begin = var_1254_begin_0, end = var_1254_end_0, end_mask = var_1254_end_mask_0, x = query_5_cast_fp16)[name = tensor<string, []>("op_1254_cast_fp16")];
907 tensor<int32, [4]> var_1258_begin_0 = const()[name = tensor<string, []>("op_1258_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
908 tensor<int32, [4]> var_1258_end_0 = const()[name = tensor<string, []>("op_1258_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
909 tensor<bool, [4]> var_1258_end_mask_0 = const()[name = tensor<string, []>("op_1258_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
910 tensor<fp16, [1, 64, 1, 1500]> var_1258_cast_fp16 = slice_by_index(begin = var_1258_begin_0, end = var_1258_end_0, end_mask = var_1258_end_mask_0, x = query_5_cast_fp16)[name = tensor<string, []>("op_1258_cast_fp16")];
911 tensor<int32, [4]> var_1262_begin_0 = const()[name = tensor<string, []>("op_1262_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
912 tensor<int32, [4]> var_1262_end_0 = const()[name = tensor<string, []>("op_1262_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
913 tensor<bool, [4]> var_1262_end_mask_0 = const()[name = tensor<string, []>("op_1262_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
914 tensor<fp16, [1, 64, 1, 1500]> var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, x = query_5_cast_fp16)[name = tensor<string, []>("op_1262_cast_fp16")];
915 tensor<int32, [4]> var_1266_begin_0 = const()[name = tensor<string, []>("op_1266_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
916 tensor<int32, [4]> var_1266_end_0 = const()[name = tensor<string, []>("op_1266_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
917 tensor<bool, [4]> var_1266_end_mask_0 = const()[name = tensor<string, []>("op_1266_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
918 tensor<fp16, [1, 64, 1, 1500]> var_1266_cast_fp16 = slice_by_index(begin = var_1266_begin_0, end = var_1266_end_0, end_mask = var_1266_end_mask_0, x = query_5_cast_fp16)[name = tensor<string, []>("op_1266_cast_fp16")];
919 tensor<int32, [4]> var_1270_begin_0 = const()[name = tensor<string, []>("op_1270_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
920 tensor<int32, [4]> var_1270_end_0 = const()[name = tensor<string, []>("op_1270_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
921 tensor<bool, [4]> var_1270_end_mask_0 = const()[name = tensor<string, []>("op_1270_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
922 tensor<fp16, [1, 64, 1, 1500]> var_1270_cast_fp16 = slice_by_index(begin = var_1270_begin_0, end = var_1270_end_0, end_mask = var_1270_end_mask_0, x = query_5_cast_fp16)[name = tensor<string, []>("op_1270_cast_fp16")];
923 tensor<int32, [4]> var_1279_begin_0 = const()[name = tensor<string, []>("op_1279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
924 tensor<int32, [4]> var_1279_end_0 = const()[name = tensor<string, []>("op_1279_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
925 tensor<bool, [4]> var_1279_end_mask_0 = const()[name = tensor<string, []>("op_1279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
926 tensor<fp16, [1, 64, 1, 375]> var_1279_cast_fp16 = slice_by_index(begin = var_1279_begin_0, end = var_1279_end_0, end_mask = var_1279_end_mask_0, x = var_1250_cast_fp16)[name = tensor<string, []>("op_1279_cast_fp16")];
927 tensor<int32, [4]> var_1286_begin_0 = const()[name = tensor<string, []>("op_1286_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
928 tensor<int32, [4]> var_1286_end_0 = const()[name = tensor<string, []>("op_1286_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
929 tensor<bool, [4]> var_1286_end_mask_0 = const()[name = tensor<string, []>("op_1286_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
930 tensor<fp16, [1, 64, 1, 375]> var_1286_cast_fp16 = slice_by_index(begin = var_1286_begin_0, end = var_1286_end_0, end_mask = var_1286_end_mask_0, x = var_1250_cast_fp16)[name = tensor<string, []>("op_1286_cast_fp16")];
931 tensor<int32, [4]> var_1293_begin_0 = const()[name = tensor<string, []>("op_1293_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
932 tensor<int32, [4]> var_1293_end_0 = const()[name = tensor<string, []>("op_1293_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
933 tensor<bool, [4]> var_1293_end_mask_0 = const()[name = tensor<string, []>("op_1293_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
934 tensor<fp16, [1, 64, 1, 375]> var_1293_cast_fp16 = slice_by_index(begin = var_1293_begin_0, end = var_1293_end_0, end_mask = var_1293_end_mask_0, x = var_1250_cast_fp16)[name = tensor<string, []>("op_1293_cast_fp16")];
935 tensor<int32, [4]> var_1300_begin_0 = const()[name = tensor<string, []>("op_1300_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
936 tensor<int32, [4]> var_1300_end_0 = const()[name = tensor<string, []>("op_1300_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
937 tensor<bool, [4]> var_1300_end_mask_0 = const()[name = tensor<string, []>("op_1300_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
938 tensor<fp16, [1, 64, 1, 375]> var_1300_cast_fp16 = slice_by_index(begin = var_1300_begin_0, end = var_1300_end_0, end_mask = var_1300_end_mask_0, x = var_1250_cast_fp16)[name = tensor<string, []>("op_1300_cast_fp16")];
939 tensor<int32, [4]> var_1307_begin_0 = const()[name = tensor<string, []>("op_1307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
940 tensor<int32, [4]> var_1307_end_0 = const()[name = tensor<string, []>("op_1307_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
941 tensor<bool, [4]> var_1307_end_mask_0 = const()[name = tensor<string, []>("op_1307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
942 tensor<fp16, [1, 64, 1, 375]> var_1307_cast_fp16 = slice_by_index(begin = var_1307_begin_0, end = var_1307_end_0, end_mask = var_1307_end_mask_0, x = var_1254_cast_fp16)[name = tensor<string, []>("op_1307_cast_fp16")];
943 tensor<int32, [4]> var_1314_begin_0 = const()[name = tensor<string, []>("op_1314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
944 tensor<int32, [4]> var_1314_end_0 = const()[name = tensor<string, []>("op_1314_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
945 tensor<bool, [4]> var_1314_end_mask_0 = const()[name = tensor<string, []>("op_1314_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
946 tensor<fp16, [1, 64, 1, 375]> var_1314_cast_fp16 = slice_by_index(begin = var_1314_begin_0, end = var_1314_end_0, end_mask = var_1314_end_mask_0, x = var_1254_cast_fp16)[name = tensor<string, []>("op_1314_cast_fp16")];
947 tensor<int32, [4]> var_1321_begin_0 = const()[name = tensor<string, []>("op_1321_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
948 tensor<int32, [4]> var_1321_end_0 = const()[name = tensor<string, []>("op_1321_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
949 tensor<bool, [4]> var_1321_end_mask_0 = const()[name = tensor<string, []>("op_1321_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
950 tensor<fp16, [1, 64, 1, 375]> var_1321_cast_fp16 = slice_by_index(begin = var_1321_begin_0, end = var_1321_end_0, end_mask = var_1321_end_mask_0, x = var_1254_cast_fp16)[name = tensor<string, []>("op_1321_cast_fp16")];
951 tensor<int32, [4]> var_1328_begin_0 = const()[name = tensor<string, []>("op_1328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
952 tensor<int32, [4]> var_1328_end_0 = const()[name = tensor<string, []>("op_1328_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
953 tensor<bool, [4]> var_1328_end_mask_0 = const()[name = tensor<string, []>("op_1328_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
954 tensor<fp16, [1, 64, 1, 375]> var_1328_cast_fp16 = slice_by_index(begin = var_1328_begin_0, end = var_1328_end_0, end_mask = var_1328_end_mask_0, x = var_1254_cast_fp16)[name = tensor<string, []>("op_1328_cast_fp16")];
955 tensor<int32, [4]> var_1335_begin_0 = const()[name = tensor<string, []>("op_1335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
956 tensor<int32, [4]> var_1335_end_0 = const()[name = tensor<string, []>("op_1335_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
957 tensor<bool, [4]> var_1335_end_mask_0 = const()[name = tensor<string, []>("op_1335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
958 tensor<fp16, [1, 64, 1, 375]> var_1335_cast_fp16 = slice_by_index(begin = var_1335_begin_0, end = var_1335_end_0, end_mask = var_1335_end_mask_0, x = var_1258_cast_fp16)[name = tensor<string, []>("op_1335_cast_fp16")];
959 tensor<int32, [4]> var_1342_begin_0 = const()[name = tensor<string, []>("op_1342_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
960 tensor<int32, [4]> var_1342_end_0 = const()[name = tensor<string, []>("op_1342_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
961 tensor<bool, [4]> var_1342_end_mask_0 = const()[name = tensor<string, []>("op_1342_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
962 tensor<fp16, [1, 64, 1, 375]> var_1342_cast_fp16 = slice_by_index(begin = var_1342_begin_0, end = var_1342_end_0, end_mask = var_1342_end_mask_0, x = var_1258_cast_fp16)[name = tensor<string, []>("op_1342_cast_fp16")];
963 tensor<int32, [4]> var_1349_begin_0 = const()[name = tensor<string, []>("op_1349_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
964 tensor<int32, [4]> var_1349_end_0 = const()[name = tensor<string, []>("op_1349_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
965 tensor<bool, [4]> var_1349_end_mask_0 = const()[name = tensor<string, []>("op_1349_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
966 tensor<fp16, [1, 64, 1, 375]> var_1349_cast_fp16 = slice_by_index(begin = var_1349_begin_0, end = var_1349_end_0, end_mask = var_1349_end_mask_0, x = var_1258_cast_fp16)[name = tensor<string, []>("op_1349_cast_fp16")];
967 tensor<int32, [4]> var_1356_begin_0 = const()[name = tensor<string, []>("op_1356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
968 tensor<int32, [4]> var_1356_end_0 = const()[name = tensor<string, []>("op_1356_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
969 tensor<bool, [4]> var_1356_end_mask_0 = const()[name = tensor<string, []>("op_1356_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
970 tensor<fp16, [1, 64, 1, 375]> var_1356_cast_fp16 = slice_by_index(begin = var_1356_begin_0, end = var_1356_end_0, end_mask = var_1356_end_mask_0, x = var_1258_cast_fp16)[name = tensor<string, []>("op_1356_cast_fp16")];
971 tensor<int32, [4]> var_1363_begin_0 = const()[name = tensor<string, []>("op_1363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
972 tensor<int32, [4]> var_1363_end_0 = const()[name = tensor<string, []>("op_1363_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
973 tensor<bool, [4]> var_1363_end_mask_0 = const()[name = tensor<string, []>("op_1363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
974 tensor<fp16, [1, 64, 1, 375]> var_1363_cast_fp16 = slice_by_index(begin = var_1363_begin_0, end = var_1363_end_0, end_mask = var_1363_end_mask_0, x = var_1262_cast_fp16)[name = tensor<string, []>("op_1363_cast_fp16")];
975 tensor<int32, [4]> var_1370_begin_0 = const()[name = tensor<string, []>("op_1370_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
976 tensor<int32, [4]> var_1370_end_0 = const()[name = tensor<string, []>("op_1370_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
977 tensor<bool, [4]> var_1370_end_mask_0 = const()[name = tensor<string, []>("op_1370_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
978 tensor<fp16, [1, 64, 1, 375]> var_1370_cast_fp16 = slice_by_index(begin = var_1370_begin_0, end = var_1370_end_0, end_mask = var_1370_end_mask_0, x = var_1262_cast_fp16)[name = tensor<string, []>("op_1370_cast_fp16")];
979 tensor<int32, [4]> var_1377_begin_0 = const()[name = tensor<string, []>("op_1377_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
980 tensor<int32, [4]> var_1377_end_0 = const()[name = tensor<string, []>("op_1377_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
981 tensor<bool, [4]> var_1377_end_mask_0 = const()[name = tensor<string, []>("op_1377_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
982 tensor<fp16, [1, 64, 1, 375]> var_1377_cast_fp16 = slice_by_index(begin = var_1377_begin_0, end = var_1377_end_0, end_mask = var_1377_end_mask_0, x = var_1262_cast_fp16)[name = tensor<string, []>("op_1377_cast_fp16")];
983 tensor<int32, [4]> var_1384_begin_0 = const()[name = tensor<string, []>("op_1384_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
984 tensor<int32, [4]> var_1384_end_0 = const()[name = tensor<string, []>("op_1384_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
985 tensor<bool, [4]> var_1384_end_mask_0 = const()[name = tensor<string, []>("op_1384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
986 tensor<fp16, [1, 64, 1, 375]> var_1384_cast_fp16 = slice_by_index(begin = var_1384_begin_0, end = var_1384_end_0, end_mask = var_1384_end_mask_0, x = var_1262_cast_fp16)[name = tensor<string, []>("op_1384_cast_fp16")];
987 tensor<int32, [4]> var_1391_begin_0 = const()[name = tensor<string, []>("op_1391_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
988 tensor<int32, [4]> var_1391_end_0 = const()[name = tensor<string, []>("op_1391_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
989 tensor<bool, [4]> var_1391_end_mask_0 = const()[name = tensor<string, []>("op_1391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
990 tensor<fp16, [1, 64, 1, 375]> var_1391_cast_fp16 = slice_by_index(begin = var_1391_begin_0, end = var_1391_end_0, end_mask = var_1391_end_mask_0, x = var_1266_cast_fp16)[name = tensor<string, []>("op_1391_cast_fp16")];
991 tensor<int32, [4]> var_1398_begin_0 = const()[name = tensor<string, []>("op_1398_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
992 tensor<int32, [4]> var_1398_end_0 = const()[name = tensor<string, []>("op_1398_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
993 tensor<bool, [4]> var_1398_end_mask_0 = const()[name = tensor<string, []>("op_1398_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
994 tensor<fp16, [1, 64, 1, 375]> var_1398_cast_fp16 = slice_by_index(begin = var_1398_begin_0, end = var_1398_end_0, end_mask = var_1398_end_mask_0, x = var_1266_cast_fp16)[name = tensor<string, []>("op_1398_cast_fp16")];
995 tensor<int32, [4]> var_1405_begin_0 = const()[name = tensor<string, []>("op_1405_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
996 tensor<int32, [4]> var_1405_end_0 = const()[name = tensor<string, []>("op_1405_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
997 tensor<bool, [4]> var_1405_end_mask_0 = const()[name = tensor<string, []>("op_1405_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
998 tensor<fp16, [1, 64, 1, 375]> var_1405_cast_fp16 = slice_by_index(begin = var_1405_begin_0, end = var_1405_end_0, end_mask = var_1405_end_mask_0, x = var_1266_cast_fp16)[name = tensor<string, []>("op_1405_cast_fp16")];
999 tensor<int32, [4]> var_1412_begin_0 = const()[name = tensor<string, []>("op_1412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1000 tensor<int32, [4]> var_1412_end_0 = const()[name = tensor<string, []>("op_1412_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1001 tensor<bool, [4]> var_1412_end_mask_0 = const()[name = tensor<string, []>("op_1412_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1002 tensor<fp16, [1, 64, 1, 375]> var_1412_cast_fp16 = slice_by_index(begin = var_1412_begin_0, end = var_1412_end_0, end_mask = var_1412_end_mask_0, x = var_1266_cast_fp16)[name = tensor<string, []>("op_1412_cast_fp16")];
1003 tensor<int32, [4]> var_1419_begin_0 = const()[name = tensor<string, []>("op_1419_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1004 tensor<int32, [4]> var_1419_end_0 = const()[name = tensor<string, []>("op_1419_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1005 tensor<bool, [4]> var_1419_end_mask_0 = const()[name = tensor<string, []>("op_1419_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1006 tensor<fp16, [1, 64, 1, 375]> var_1419_cast_fp16 = slice_by_index(begin = var_1419_begin_0, end = var_1419_end_0, end_mask = var_1419_end_mask_0, x = var_1270_cast_fp16)[name = tensor<string, []>("op_1419_cast_fp16")];
1007 tensor<int32, [4]> var_1426_begin_0 = const()[name = tensor<string, []>("op_1426_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1008 tensor<int32, [4]> var_1426_end_0 = const()[name = tensor<string, []>("op_1426_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1009 tensor<bool, [4]> var_1426_end_mask_0 = const()[name = tensor<string, []>("op_1426_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1010 tensor<fp16, [1, 64, 1, 375]> var_1426_cast_fp16 = slice_by_index(begin = var_1426_begin_0, end = var_1426_end_0, end_mask = var_1426_end_mask_0, x = var_1270_cast_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
1011 tensor<int32, [4]> var_1433_begin_0 = const()[name = tensor<string, []>("op_1433_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1012 tensor<int32, [4]> var_1433_end_0 = const()[name = tensor<string, []>("op_1433_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1013 tensor<bool, [4]> var_1433_end_mask_0 = const()[name = tensor<string, []>("op_1433_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1014 tensor<fp16, [1, 64, 1, 375]> var_1433_cast_fp16 = slice_by_index(begin = var_1433_begin_0, end = var_1433_end_0, end_mask = var_1433_end_mask_0, x = var_1270_cast_fp16)[name = tensor<string, []>("op_1433_cast_fp16")];
1015 tensor<int32, [4]> var_1440_begin_0 = const()[name = tensor<string, []>("op_1440_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1016 tensor<int32, [4]> var_1440_end_0 = const()[name = tensor<string, []>("op_1440_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1017 tensor<bool, [4]> var_1440_end_mask_0 = const()[name = tensor<string, []>("op_1440_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1018 tensor<fp16, [1, 64, 1, 375]> var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = var_1270_cast_fp16)[name = tensor<string, []>("op_1440_cast_fp16")];
1019 tensor<int32, [4]> k_5_perm_0 = const()[name = tensor<string, []>("k_5_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
1020 tensor<int32, [4]> var_1445_begin_0 = const()[name = tensor<string, []>("op_1445_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1021 tensor<int32, [4]> var_1445_end_0 = const()[name = tensor<string, []>("op_1445_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 64])];
1022 tensor<bool, [4]> var_1445_end_mask_0 = const()[name = tensor<string, []>("op_1445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1023 tensor<fp16, [1, 1500, 1, 384]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = key_5_cast_fp16)[name = tensor<string, []>("transpose_1")];
1024 tensor<fp16, [1, 1500, 1, 64]> var_1445_cast_fp16 = slice_by_index(begin = var_1445_begin_0, end = var_1445_end_0, end_mask = var_1445_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("op_1445_cast_fp16")];
1025 tensor<int32, [4]> var_1449_begin_0 = const()[name = tensor<string, []>("op_1449_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
1026 tensor<int32, [4]> var_1449_end_0 = const()[name = tensor<string, []>("op_1449_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 128])];
1027 tensor<bool, [4]> var_1449_end_mask_0 = const()[name = tensor<string, []>("op_1449_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1028 tensor<fp16, [1, 1500, 1, 64]> var_1449_cast_fp16 = slice_by_index(begin = var_1449_begin_0, end = var_1449_end_0, end_mask = var_1449_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("op_1449_cast_fp16")];
1029 tensor<int32, [4]> var_1453_begin_0 = const()[name = tensor<string, []>("op_1453_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
1030 tensor<int32, [4]> var_1453_end_0 = const()[name = tensor<string, []>("op_1453_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 192])];
1031 tensor<bool, [4]> var_1453_end_mask_0 = const()[name = tensor<string, []>("op_1453_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1032 tensor<fp16, [1, 1500, 1, 64]> var_1453_cast_fp16 = slice_by_index(begin = var_1453_begin_0, end = var_1453_end_0, end_mask = var_1453_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("op_1453_cast_fp16")];
1033 tensor<int32, [4]> var_1457_begin_0 = const()[name = tensor<string, []>("op_1457_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 192])];
1034 tensor<int32, [4]> var_1457_end_0 = const()[name = tensor<string, []>("op_1457_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 256])];
1035 tensor<bool, [4]> var_1457_end_mask_0 = const()[name = tensor<string, []>("op_1457_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1036 tensor<fp16, [1, 1500, 1, 64]> var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("op_1457_cast_fp16")];
1037 tensor<int32, [4]> var_1461_begin_0 = const()[name = tensor<string, []>("op_1461_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
1038 tensor<int32, [4]> var_1461_end_0 = const()[name = tensor<string, []>("op_1461_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 320])];
1039 tensor<bool, [4]> var_1461_end_mask_0 = const()[name = tensor<string, []>("op_1461_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1040 tensor<fp16, [1, 1500, 1, 64]> var_1461_cast_fp16 = slice_by_index(begin = var_1461_begin_0, end = var_1461_end_0, end_mask = var_1461_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("op_1461_cast_fp16")];
1041 tensor<int32, [4]> var_1465_begin_0 = const()[name = tensor<string, []>("op_1465_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 320])];
1042 tensor<int32, [4]> var_1465_end_0 = const()[name = tensor<string, []>("op_1465_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 384])];
1043 tensor<bool, [4]> var_1465_end_mask_0 = const()[name = tensor<string, []>("op_1465_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1044 tensor<fp16, [1, 1500, 1, 64]> var_1465_cast_fp16 = slice_by_index(begin = var_1465_begin_0, end = var_1465_end_0, end_mask = var_1465_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("op_1465_cast_fp16")];
1045 tensor<int32, [4]> var_1467_begin_0 = const()[name = tensor<string, []>("op_1467_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1046 tensor<int32, [4]> var_1467_end_0 = const()[name = tensor<string, []>("op_1467_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1047 tensor<bool, [4]> var_1467_end_mask_0 = const()[name = tensor<string, []>("op_1467_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1048 tensor<fp16, [1, 64, 1, 1500]> var_1467_cast_fp16 = slice_by_index(begin = var_1467_begin_0, end = var_1467_end_0, end_mask = var_1467_end_mask_0, x = value_5_cast_fp16)[name = tensor<string, []>("op_1467_cast_fp16")];
1049 tensor<int32, [4]> var_1471_begin_0 = const()[name = tensor<string, []>("op_1471_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
1050 tensor<int32, [4]> var_1471_end_0 = const()[name = tensor<string, []>("op_1471_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
1051 tensor<bool, [4]> var_1471_end_mask_0 = const()[name = tensor<string, []>("op_1471_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1052 tensor<fp16, [1, 64, 1, 1500]> var_1471_cast_fp16 = slice_by_index(begin = var_1471_begin_0, end = var_1471_end_0, end_mask = var_1471_end_mask_0, x = value_5_cast_fp16)[name = tensor<string, []>("op_1471_cast_fp16")];
1053 tensor<int32, [4]> var_1475_begin_0 = const()[name = tensor<string, []>("op_1475_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
1054 tensor<int32, [4]> var_1475_end_0 = const()[name = tensor<string, []>("op_1475_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
1055 tensor<bool, [4]> var_1475_end_mask_0 = const()[name = tensor<string, []>("op_1475_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1056 tensor<fp16, [1, 64, 1, 1500]> var_1475_cast_fp16 = slice_by_index(begin = var_1475_begin_0, end = var_1475_end_0, end_mask = var_1475_end_mask_0, x = value_5_cast_fp16)[name = tensor<string, []>("op_1475_cast_fp16")];
1057 tensor<int32, [4]> var_1479_begin_0 = const()[name = tensor<string, []>("op_1479_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
1058 tensor<int32, [4]> var_1479_end_0 = const()[name = tensor<string, []>("op_1479_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
1059 tensor<bool, [4]> var_1479_end_mask_0 = const()[name = tensor<string, []>("op_1479_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1060 tensor<fp16, [1, 64, 1, 1500]> var_1479_cast_fp16 = slice_by_index(begin = var_1479_begin_0, end = var_1479_end_0, end_mask = var_1479_end_mask_0, x = value_5_cast_fp16)[name = tensor<string, []>("op_1479_cast_fp16")];
1061 tensor<int32, [4]> var_1483_begin_0 = const()[name = tensor<string, []>("op_1483_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
1062 tensor<int32, [4]> var_1483_end_0 = const()[name = tensor<string, []>("op_1483_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
1063 tensor<bool, [4]> var_1483_end_mask_0 = const()[name = tensor<string, []>("op_1483_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1064 tensor<fp16, [1, 64, 1, 1500]> var_1483_cast_fp16 = slice_by_index(begin = var_1483_begin_0, end = var_1483_end_0, end_mask = var_1483_end_mask_0, x = value_5_cast_fp16)[name = tensor<string, []>("op_1483_cast_fp16")];
1065 tensor<int32, [4]> var_1487_begin_0 = const()[name = tensor<string, []>("op_1487_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
1066 tensor<int32, [4]> var_1487_end_0 = const()[name = tensor<string, []>("op_1487_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
1067 tensor<bool, [4]> var_1487_end_mask_0 = const()[name = tensor<string, []>("op_1487_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1068 tensor<fp16, [1, 64, 1, 1500]> var_1487_cast_fp16 = slice_by_index(begin = var_1487_begin_0, end = var_1487_end_0, end_mask = var_1487_end_mask_0, x = value_5_cast_fp16)[name = tensor<string, []>("op_1487_cast_fp16")];
1069 tensor<string, []> _SplitHeadsQ__mh_w_97_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_97_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1070 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_97_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_97_equation_0, values = (var_1445_cast_fp16, var_1279_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_97_cast_fp16")];
1071 tensor<string, []> _SplitHeadsQ__mh_w_99_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_99_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1072 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_99_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_99_equation_0, values = (var_1445_cast_fp16, var_1286_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_99_cast_fp16")];
1073 tensor<string, []> _SplitHeadsQ__mh_w_101_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_101_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1074 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_101_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_101_equation_0, values = (var_1445_cast_fp16, var_1293_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_101_cast_fp16")];
1075 tensor<string, []> _SplitHeadsQ__mh_w_103_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_103_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1076 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_103_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_103_equation_0, values = (var_1445_cast_fp16, var_1300_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_103_cast_fp16")];
1077 tensor<string, []> _SplitHeadsQ__mh_w_105_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_105_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1078 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_105_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_105_equation_0, values = (var_1449_cast_fp16, var_1307_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_105_cast_fp16")];
1079 tensor<string, []> _SplitHeadsQ__mh_w_107_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_107_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1080 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_107_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_107_equation_0, values = (var_1449_cast_fp16, var_1314_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_107_cast_fp16")];
1081 tensor<string, []> _SplitHeadsQ__mh_w_109_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_109_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1082 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_109_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_109_equation_0, values = (var_1449_cast_fp16, var_1321_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_109_cast_fp16")];
1083 tensor<string, []> _SplitHeadsQ__mh_w_111_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_111_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1084 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_111_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_111_equation_0, values = (var_1449_cast_fp16, var_1328_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_111_cast_fp16")];
1085 tensor<string, []> _SplitHeadsQ__mh_w_113_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_113_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1086 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_113_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_113_equation_0, values = (var_1453_cast_fp16, var_1335_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_113_cast_fp16")];
1087 tensor<string, []> _SplitHeadsQ__mh_w_115_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_115_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1088 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_115_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_115_equation_0, values = (var_1453_cast_fp16, var_1342_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_115_cast_fp16")];
1089 tensor<string, []> _SplitHeadsQ__mh_w_117_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_117_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1090 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_117_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_117_equation_0, values = (var_1453_cast_fp16, var_1349_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_117_cast_fp16")];
1091 tensor<string, []> _SplitHeadsQ__mh_w_119_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_119_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1092 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_119_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_119_equation_0, values = (var_1453_cast_fp16, var_1356_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_119_cast_fp16")];
1093 tensor<string, []> _SplitHeadsQ__mh_w_121_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_121_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1094 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_121_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_121_equation_0, values = (var_1457_cast_fp16, var_1363_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_121_cast_fp16")];
1095 tensor<string, []> _SplitHeadsQ__mh_w_123_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_123_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1096 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_123_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_123_equation_0, values = (var_1457_cast_fp16, var_1370_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_123_cast_fp16")];
1097 tensor<string, []> _SplitHeadsQ__mh_w_125_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_125_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1098 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_125_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_125_equation_0, values = (var_1457_cast_fp16, var_1377_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_125_cast_fp16")];
1099 tensor<string, []> _SplitHeadsQ__mh_w_127_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_127_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1100 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_127_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_127_equation_0, values = (var_1457_cast_fp16, var_1384_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_127_cast_fp16")];
1101 tensor<string, []> _SplitHeadsQ__mh_w_129_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_129_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1102 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_129_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_129_equation_0, values = (var_1461_cast_fp16, var_1391_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_129_cast_fp16")];
1103 tensor<string, []> _SplitHeadsQ__mh_w_131_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_131_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1104 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_131_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_131_equation_0, values = (var_1461_cast_fp16, var_1398_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_131_cast_fp16")];
1105 tensor<string, []> _SplitHeadsQ__mh_w_133_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_133_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1106 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_133_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_133_equation_0, values = (var_1461_cast_fp16, var_1405_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_133_cast_fp16")];
1107 tensor<string, []> _SplitHeadsQ__mh_w_135_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_135_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1108 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_135_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_135_equation_0, values = (var_1461_cast_fp16, var_1412_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_135_cast_fp16")];
1109 tensor<string, []> _SplitHeadsQ__mh_w_137_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_137_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1110 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_137_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_137_equation_0, values = (var_1465_cast_fp16, var_1419_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_137_cast_fp16")];
1111 tensor<string, []> _SplitHeadsQ__mh_w_139_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_139_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1112 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_139_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_139_equation_0, values = (var_1465_cast_fp16, var_1426_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_139_cast_fp16")];
1113 tensor<string, []> _SplitHeadsQ__mh_w_141_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_141_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1114 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_141_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_141_equation_0, values = (var_1465_cast_fp16, var_1433_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_141_cast_fp16")];
1115 tensor<string, []> _SplitHeadsQ__mh_w_143_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_143_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1116 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_143_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_143_equation_0, values = (var_1465_cast_fp16, var_1440_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_143_cast_fp16")];
1117 tensor<fp16, []> var_1538_to_fp16 = const()[name = tensor<string, []>("op_1538_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1118 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_97_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_97_cast_fp16, y = var_1538_to_fp16)[name = tensor<string, []>("aw_chunk_97_cast_fp16")];
1119 tensor<fp16, []> var_1540_to_fp16 = const()[name = tensor<string, []>("op_1540_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1120 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_99_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_99_cast_fp16, y = var_1540_to_fp16)[name = tensor<string, []>("aw_chunk_99_cast_fp16")];
1121 tensor<fp16, []> var_1542_to_fp16 = const()[name = tensor<string, []>("op_1542_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1122 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_101_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_101_cast_fp16, y = var_1542_to_fp16)[name = tensor<string, []>("aw_chunk_101_cast_fp16")];
1123 tensor<fp16, []> var_1544_to_fp16 = const()[name = tensor<string, []>("op_1544_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1124 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_103_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_103_cast_fp16, y = var_1544_to_fp16)[name = tensor<string, []>("aw_chunk_103_cast_fp16")];
1125 tensor<fp16, []> var_1546_to_fp16 = const()[name = tensor<string, []>("op_1546_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1126 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_105_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_105_cast_fp16, y = var_1546_to_fp16)[name = tensor<string, []>("aw_chunk_105_cast_fp16")];
1127 tensor<fp16, []> var_1548_to_fp16 = const()[name = tensor<string, []>("op_1548_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1128 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_107_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_107_cast_fp16, y = var_1548_to_fp16)[name = tensor<string, []>("aw_chunk_107_cast_fp16")];
1129 tensor<fp16, []> var_1550_to_fp16 = const()[name = tensor<string, []>("op_1550_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1130 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_109_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_109_cast_fp16, y = var_1550_to_fp16)[name = tensor<string, []>("aw_chunk_109_cast_fp16")];
1131 tensor<fp16, []> var_1552_to_fp16 = const()[name = tensor<string, []>("op_1552_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1132 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_111_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_111_cast_fp16, y = var_1552_to_fp16)[name = tensor<string, []>("aw_chunk_111_cast_fp16")];
1133 tensor<fp16, []> var_1554_to_fp16 = const()[name = tensor<string, []>("op_1554_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1134 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_113_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_113_cast_fp16, y = var_1554_to_fp16)[name = tensor<string, []>("aw_chunk_113_cast_fp16")];
1135 tensor<fp16, []> var_1556_to_fp16 = const()[name = tensor<string, []>("op_1556_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1136 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_115_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_115_cast_fp16, y = var_1556_to_fp16)[name = tensor<string, []>("aw_chunk_115_cast_fp16")];
1137 tensor<fp16, []> var_1558_to_fp16 = const()[name = tensor<string, []>("op_1558_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1138 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_117_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_117_cast_fp16, y = var_1558_to_fp16)[name = tensor<string, []>("aw_chunk_117_cast_fp16")];
1139 tensor<fp16, []> var_1560_to_fp16 = const()[name = tensor<string, []>("op_1560_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1140 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_119_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_119_cast_fp16, y = var_1560_to_fp16)[name = tensor<string, []>("aw_chunk_119_cast_fp16")];
1141 tensor<fp16, []> var_1562_to_fp16 = const()[name = tensor<string, []>("op_1562_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1142 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_121_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_121_cast_fp16, y = var_1562_to_fp16)[name = tensor<string, []>("aw_chunk_121_cast_fp16")];
1143 tensor<fp16, []> var_1564_to_fp16 = const()[name = tensor<string, []>("op_1564_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1144 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_123_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_123_cast_fp16, y = var_1564_to_fp16)[name = tensor<string, []>("aw_chunk_123_cast_fp16")];
1145 tensor<fp16, []> var_1566_to_fp16 = const()[name = tensor<string, []>("op_1566_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1146 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_125_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_125_cast_fp16, y = var_1566_to_fp16)[name = tensor<string, []>("aw_chunk_125_cast_fp16")];
1147 tensor<fp16, []> var_1568_to_fp16 = const()[name = tensor<string, []>("op_1568_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1148 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_127_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_127_cast_fp16, y = var_1568_to_fp16)[name = tensor<string, []>("aw_chunk_127_cast_fp16")];
1149 tensor<fp16, []> var_1570_to_fp16 = const()[name = tensor<string, []>("op_1570_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1150 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_129_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_129_cast_fp16, y = var_1570_to_fp16)[name = tensor<string, []>("aw_chunk_129_cast_fp16")];
1151 tensor<fp16, []> var_1572_to_fp16 = const()[name = tensor<string, []>("op_1572_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1152 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_131_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_131_cast_fp16, y = var_1572_to_fp16)[name = tensor<string, []>("aw_chunk_131_cast_fp16")];
1153 tensor<fp16, []> var_1574_to_fp16 = const()[name = tensor<string, []>("op_1574_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1154 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_133_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_133_cast_fp16, y = var_1574_to_fp16)[name = tensor<string, []>("aw_chunk_133_cast_fp16")];
1155 tensor<fp16, []> var_1576_to_fp16 = const()[name = tensor<string, []>("op_1576_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1156 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_135_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_135_cast_fp16, y = var_1576_to_fp16)[name = tensor<string, []>("aw_chunk_135_cast_fp16")];
1157 tensor<fp16, []> var_1578_to_fp16 = const()[name = tensor<string, []>("op_1578_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1158 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_137_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_137_cast_fp16, y = var_1578_to_fp16)[name = tensor<string, []>("aw_chunk_137_cast_fp16")];
1159 tensor<fp16, []> var_1580_to_fp16 = const()[name = tensor<string, []>("op_1580_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1160 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_139_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_139_cast_fp16, y = var_1580_to_fp16)[name = tensor<string, []>("aw_chunk_139_cast_fp16")];
1161 tensor<fp16, []> var_1582_to_fp16 = const()[name = tensor<string, []>("op_1582_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1162 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_141_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_141_cast_fp16, y = var_1582_to_fp16)[name = tensor<string, []>("aw_chunk_141_cast_fp16")];
1163 tensor<fp16, []> var_1584_to_fp16 = const()[name = tensor<string, []>("op_1584_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1164 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_143_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_143_cast_fp16, y = var_1584_to_fp16)[name = tensor<string, []>("aw_chunk_143_cast_fp16")];
1165 tensor<fp16, [1, 1500, 1, 375]> var_1586_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_97_cast_fp16)[name = tensor<string, []>("op_1586_cast_fp16")];
1166 tensor<fp16, [1, 1500, 1, 375]> var_1587_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_99_cast_fp16)[name = tensor<string, []>("op_1587_cast_fp16")];
1167 tensor<fp16, [1, 1500, 1, 375]> var_1588_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_101_cast_fp16)[name = tensor<string, []>("op_1588_cast_fp16")];
1168 tensor<fp16, [1, 1500, 1, 375]> var_1589_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_103_cast_fp16)[name = tensor<string, []>("op_1589_cast_fp16")];
1169 tensor<fp16, [1, 1500, 1, 375]> var_1590_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_105_cast_fp16)[name = tensor<string, []>("op_1590_cast_fp16")];
1170 tensor<fp16, [1, 1500, 1, 375]> var_1591_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_107_cast_fp16)[name = tensor<string, []>("op_1591_cast_fp16")];
1171 tensor<fp16, [1, 1500, 1, 375]> var_1592_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_109_cast_fp16)[name = tensor<string, []>("op_1592_cast_fp16")];
1172 tensor<fp16, [1, 1500, 1, 375]> var_1593_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_111_cast_fp16)[name = tensor<string, []>("op_1593_cast_fp16")];
1173 tensor<fp16, [1, 1500, 1, 375]> var_1594_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_113_cast_fp16)[name = tensor<string, []>("op_1594_cast_fp16")];
1174 tensor<fp16, [1, 1500, 1, 375]> var_1595_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_115_cast_fp16)[name = tensor<string, []>("op_1595_cast_fp16")];
1175 tensor<fp16, [1, 1500, 1, 375]> var_1596_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_117_cast_fp16)[name = tensor<string, []>("op_1596_cast_fp16")];
1176 tensor<fp16, [1, 1500, 1, 375]> var_1597_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_119_cast_fp16)[name = tensor<string, []>("op_1597_cast_fp16")];
1177 tensor<fp16, [1, 1500, 1, 375]> var_1598_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_121_cast_fp16)[name = tensor<string, []>("op_1598_cast_fp16")];
1178 tensor<fp16, [1, 1500, 1, 375]> var_1599_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_123_cast_fp16)[name = tensor<string, []>("op_1599_cast_fp16")];
1179 tensor<fp16, [1, 1500, 1, 375]> var_1600_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_125_cast_fp16)[name = tensor<string, []>("op_1600_cast_fp16")];
1180 tensor<fp16, [1, 1500, 1, 375]> var_1601_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_127_cast_fp16)[name = tensor<string, []>("op_1601_cast_fp16")];
1181 tensor<fp16, [1, 1500, 1, 375]> var_1602_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_129_cast_fp16)[name = tensor<string, []>("op_1602_cast_fp16")];
1182 tensor<fp16, [1, 1500, 1, 375]> var_1603_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_131_cast_fp16)[name = tensor<string, []>("op_1603_cast_fp16")];
1183 tensor<fp16, [1, 1500, 1, 375]> var_1604_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_133_cast_fp16)[name = tensor<string, []>("op_1604_cast_fp16")];
1184 tensor<fp16, [1, 1500, 1, 375]> var_1605_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_135_cast_fp16)[name = tensor<string, []>("op_1605_cast_fp16")];
1185 tensor<fp16, [1, 1500, 1, 375]> var_1606_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_137_cast_fp16)[name = tensor<string, []>("op_1606_cast_fp16")];
1186 tensor<fp16, [1, 1500, 1, 375]> var_1607_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_139_cast_fp16)[name = tensor<string, []>("op_1607_cast_fp16")];
1187 tensor<fp16, [1, 1500, 1, 375]> var_1608_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_141_cast_fp16)[name = tensor<string, []>("op_1608_cast_fp16")];
1188 tensor<fp16, [1, 1500, 1, 375]> var_1609_cast_fp16 = softmax(axis = var_1195, x = aw_chunk_143_cast_fp16)[name = tensor<string, []>("op_1609_cast_fp16")];
1189 tensor<string, []> var_1611_equation_0 = const()[name = tensor<string, []>("op_1611_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1190 tensor<fp16, [1, 64, 1, 375]> var_1611_cast_fp16 = einsum(equation = var_1611_equation_0, values = (var_1467_cast_fp16, var_1586_cast_fp16))[name = tensor<string, []>("op_1611_cast_fp16")];
1191 tensor<string, []> var_1613_equation_0 = const()[name = tensor<string, []>("op_1613_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1192 tensor<fp16, [1, 64, 1, 375]> var_1613_cast_fp16 = einsum(equation = var_1613_equation_0, values = (var_1467_cast_fp16, var_1587_cast_fp16))[name = tensor<string, []>("op_1613_cast_fp16")];
1193 tensor<string, []> var_1615_equation_0 = const()[name = tensor<string, []>("op_1615_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1194 tensor<fp16, [1, 64, 1, 375]> var_1615_cast_fp16 = einsum(equation = var_1615_equation_0, values = (var_1467_cast_fp16, var_1588_cast_fp16))[name = tensor<string, []>("op_1615_cast_fp16")];
1195 tensor<string, []> var_1617_equation_0 = const()[name = tensor<string, []>("op_1617_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1196 tensor<fp16, [1, 64, 1, 375]> var_1617_cast_fp16 = einsum(equation = var_1617_equation_0, values = (var_1467_cast_fp16, var_1589_cast_fp16))[name = tensor<string, []>("op_1617_cast_fp16")];
1197 tensor<string, []> var_1619_equation_0 = const()[name = tensor<string, []>("op_1619_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1198 tensor<fp16, [1, 64, 1, 375]> var_1619_cast_fp16 = einsum(equation = var_1619_equation_0, values = (var_1471_cast_fp16, var_1590_cast_fp16))[name = tensor<string, []>("op_1619_cast_fp16")];
1199 tensor<string, []> var_1621_equation_0 = const()[name = tensor<string, []>("op_1621_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1200 tensor<fp16, [1, 64, 1, 375]> var_1621_cast_fp16 = einsum(equation = var_1621_equation_0, values = (var_1471_cast_fp16, var_1591_cast_fp16))[name = tensor<string, []>("op_1621_cast_fp16")];
1201 tensor<string, []> var_1623_equation_0 = const()[name = tensor<string, []>("op_1623_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1202 tensor<fp16, [1, 64, 1, 375]> var_1623_cast_fp16 = einsum(equation = var_1623_equation_0, values = (var_1471_cast_fp16, var_1592_cast_fp16))[name = tensor<string, []>("op_1623_cast_fp16")];
1203 tensor<string, []> var_1625_equation_0 = const()[name = tensor<string, []>("op_1625_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1204 tensor<fp16, [1, 64, 1, 375]> var_1625_cast_fp16 = einsum(equation = var_1625_equation_0, values = (var_1471_cast_fp16, var_1593_cast_fp16))[name = tensor<string, []>("op_1625_cast_fp16")];
1205 tensor<string, []> var_1627_equation_0 = const()[name = tensor<string, []>("op_1627_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1206 tensor<fp16, [1, 64, 1, 375]> var_1627_cast_fp16 = einsum(equation = var_1627_equation_0, values = (var_1475_cast_fp16, var_1594_cast_fp16))[name = tensor<string, []>("op_1627_cast_fp16")];
1207 tensor<string, []> var_1629_equation_0 = const()[name = tensor<string, []>("op_1629_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1208 tensor<fp16, [1, 64, 1, 375]> var_1629_cast_fp16 = einsum(equation = var_1629_equation_0, values = (var_1475_cast_fp16, var_1595_cast_fp16))[name = tensor<string, []>("op_1629_cast_fp16")];
1209 tensor<string, []> var_1631_equation_0 = const()[name = tensor<string, []>("op_1631_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1210 tensor<fp16, [1, 64, 1, 375]> var_1631_cast_fp16 = einsum(equation = var_1631_equation_0, values = (var_1475_cast_fp16, var_1596_cast_fp16))[name = tensor<string, []>("op_1631_cast_fp16")];
1211 tensor<string, []> var_1633_equation_0 = const()[name = tensor<string, []>("op_1633_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1212 tensor<fp16, [1, 64, 1, 375]> var_1633_cast_fp16 = einsum(equation = var_1633_equation_0, values = (var_1475_cast_fp16, var_1597_cast_fp16))[name = tensor<string, []>("op_1633_cast_fp16")];
1213 tensor<string, []> var_1635_equation_0 = const()[name = tensor<string, []>("op_1635_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1214 tensor<fp16, [1, 64, 1, 375]> var_1635_cast_fp16 = einsum(equation = var_1635_equation_0, values = (var_1479_cast_fp16, var_1598_cast_fp16))[name = tensor<string, []>("op_1635_cast_fp16")];
1215 tensor<string, []> var_1637_equation_0 = const()[name = tensor<string, []>("op_1637_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1216 tensor<fp16, [1, 64, 1, 375]> var_1637_cast_fp16 = einsum(equation = var_1637_equation_0, values = (var_1479_cast_fp16, var_1599_cast_fp16))[name = tensor<string, []>("op_1637_cast_fp16")];
1217 tensor<string, []> var_1639_equation_0 = const()[name = tensor<string, []>("op_1639_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1218 tensor<fp16, [1, 64, 1, 375]> var_1639_cast_fp16 = einsum(equation = var_1639_equation_0, values = (var_1479_cast_fp16, var_1600_cast_fp16))[name = tensor<string, []>("op_1639_cast_fp16")];
1219 tensor<string, []> var_1641_equation_0 = const()[name = tensor<string, []>("op_1641_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1220 tensor<fp16, [1, 64, 1, 375]> var_1641_cast_fp16 = einsum(equation = var_1641_equation_0, values = (var_1479_cast_fp16, var_1601_cast_fp16))[name = tensor<string, []>("op_1641_cast_fp16")];
1221 tensor<string, []> var_1643_equation_0 = const()[name = tensor<string, []>("op_1643_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1222 tensor<fp16, [1, 64, 1, 375]> var_1643_cast_fp16 = einsum(equation = var_1643_equation_0, values = (var_1483_cast_fp16, var_1602_cast_fp16))[name = tensor<string, []>("op_1643_cast_fp16")];
1223 tensor<string, []> var_1645_equation_0 = const()[name = tensor<string, []>("op_1645_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1224 tensor<fp16, [1, 64, 1, 375]> var_1645_cast_fp16 = einsum(equation = var_1645_equation_0, values = (var_1483_cast_fp16, var_1603_cast_fp16))[name = tensor<string, []>("op_1645_cast_fp16")];
1225 tensor<string, []> var_1647_equation_0 = const()[name = tensor<string, []>("op_1647_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1226 tensor<fp16, [1, 64, 1, 375]> var_1647_cast_fp16 = einsum(equation = var_1647_equation_0, values = (var_1483_cast_fp16, var_1604_cast_fp16))[name = tensor<string, []>("op_1647_cast_fp16")];
1227 tensor<string, []> var_1649_equation_0 = const()[name = tensor<string, []>("op_1649_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1228 tensor<fp16, [1, 64, 1, 375]> var_1649_cast_fp16 = einsum(equation = var_1649_equation_0, values = (var_1483_cast_fp16, var_1605_cast_fp16))[name = tensor<string, []>("op_1649_cast_fp16")];
1229 tensor<string, []> var_1651_equation_0 = const()[name = tensor<string, []>("op_1651_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1230 tensor<fp16, [1, 64, 1, 375]> var_1651_cast_fp16 = einsum(equation = var_1651_equation_0, values = (var_1487_cast_fp16, var_1606_cast_fp16))[name = tensor<string, []>("op_1651_cast_fp16")];
1231 tensor<string, []> var_1653_equation_0 = const()[name = tensor<string, []>("op_1653_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1232 tensor<fp16, [1, 64, 1, 375]> var_1653_cast_fp16 = einsum(equation = var_1653_equation_0, values = (var_1487_cast_fp16, var_1607_cast_fp16))[name = tensor<string, []>("op_1653_cast_fp16")];
1233 tensor<string, []> var_1655_equation_0 = const()[name = tensor<string, []>("op_1655_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1234 tensor<fp16, [1, 64, 1, 375]> var_1655_cast_fp16 = einsum(equation = var_1655_equation_0, values = (var_1487_cast_fp16, var_1608_cast_fp16))[name = tensor<string, []>("op_1655_cast_fp16")];
1235 tensor<string, []> var_1657_equation_0 = const()[name = tensor<string, []>("op_1657_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1236 tensor<fp16, [1, 64, 1, 375]> var_1657_cast_fp16 = einsum(equation = var_1657_equation_0, values = (var_1487_cast_fp16, var_1609_cast_fp16))[name = tensor<string, []>("op_1657_cast_fp16")];
1237 tensor<bool, []> var_1659_interleave_0 = const()[name = tensor<string, []>("op_1659_interleave_0"), val = tensor<bool, []>(false)];
1238 tensor<fp16, [1, 64, 1, 1500]> var_1659_cast_fp16 = concat(axis = var_1184, interleave = var_1659_interleave_0, values = (var_1611_cast_fp16, var_1613_cast_fp16, var_1615_cast_fp16, var_1617_cast_fp16))[name = tensor<string, []>("op_1659_cast_fp16")];
1239 tensor<bool, []> var_1661_interleave_0 = const()[name = tensor<string, []>("op_1661_interleave_0"), val = tensor<bool, []>(false)];
1240 tensor<fp16, [1, 64, 1, 1500]> var_1661_cast_fp16 = concat(axis = var_1184, interleave = var_1661_interleave_0, values = (var_1619_cast_fp16, var_1621_cast_fp16, var_1623_cast_fp16, var_1625_cast_fp16))[name = tensor<string, []>("op_1661_cast_fp16")];
1241 tensor<bool, []> var_1663_interleave_0 = const()[name = tensor<string, []>("op_1663_interleave_0"), val = tensor<bool, []>(false)];
1242 tensor<fp16, [1, 64, 1, 1500]> var_1663_cast_fp16 = concat(axis = var_1184, interleave = var_1663_interleave_0, values = (var_1627_cast_fp16, var_1629_cast_fp16, var_1631_cast_fp16, var_1633_cast_fp16))[name = tensor<string, []>("op_1663_cast_fp16")];
1243 tensor<bool, []> var_1665_interleave_0 = const()[name = tensor<string, []>("op_1665_interleave_0"), val = tensor<bool, []>(false)];
1244 tensor<fp16, [1, 64, 1, 1500]> var_1665_cast_fp16 = concat(axis = var_1184, interleave = var_1665_interleave_0, values = (var_1635_cast_fp16, var_1637_cast_fp16, var_1639_cast_fp16, var_1641_cast_fp16))[name = tensor<string, []>("op_1665_cast_fp16")];
1245 tensor<bool, []> var_1667_interleave_0 = const()[name = tensor<string, []>("op_1667_interleave_0"), val = tensor<bool, []>(false)];
1246 tensor<fp16, [1, 64, 1, 1500]> var_1667_cast_fp16 = concat(axis = var_1184, interleave = var_1667_interleave_0, values = (var_1643_cast_fp16, var_1645_cast_fp16, var_1647_cast_fp16, var_1649_cast_fp16))[name = tensor<string, []>("op_1667_cast_fp16")];
1247 tensor<bool, []> var_1669_interleave_0 = const()[name = tensor<string, []>("op_1669_interleave_0"), val = tensor<bool, []>(false)];
1248 tensor<fp16, [1, 64, 1, 1500]> var_1669_cast_fp16 = concat(axis = var_1184, interleave = var_1669_interleave_0, values = (var_1651_cast_fp16, var_1653_cast_fp16, var_1655_cast_fp16, var_1657_cast_fp16))[name = tensor<string, []>("op_1669_cast_fp16")];
1249 tensor<bool, []> input_17_interleave_0 = const()[name = tensor<string, []>("input_17_interleave_0"), val = tensor<bool, []>(false)];
1250 tensor<fp16, [1, 384, 1, 1500]> input_17_cast_fp16 = concat(axis = var_1195, interleave = input_17_interleave_0, values = (var_1659_cast_fp16, var_1661_cast_fp16, var_1663_cast_fp16, var_1665_cast_fp16, var_1667_cast_fp16, var_1669_cast_fp16))[name = tensor<string, []>("input_17_cast_fp16")];
1251 tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("valid")];
1252 tensor<int32, [2]> obj_11_strides_0 = const()[name = tensor<string, []>("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
1253 tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1254 tensor<int32, [2]> obj_11_dilations_0 = const()[name = tensor<string, []>("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1255 tensor<int32, []> obj_11_groups_0 = const()[name = tensor<string, []>("obj_11_groups_0"), val = tensor<int32, []>(1)];
1256 tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10211136)))];
1257 tensor<fp16, [384]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506112)))];
1258 tensor<fp16, [1, 384, 1, 1500]> obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
1259 tensor<fp16, [1, 384, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
1260 tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
1261 tensor<fp16, []> var_1688_to_fp16 = const()[name = tensor<string, []>("op_1688_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1262 tensor<fp16, [1, 384, 1, 1500]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1688_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
1263 tensor<fp16, [384]> input_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_19_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506944)))];
1264 tensor<fp16, [384]> input_19_beta_0_to_fp16 = const()[name = tensor<string, []>("input_19_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10507776)))];
1265 tensor<fp16, []> input_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1266 tensor<fp16, [1, 384, 1, 1500]> input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_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_19_cast_fp16")];
1267 tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("valid")];
1268 tensor<int32, [2]> input_21_strides_0 = const()[name = tensor<string, []>("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
1269 tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1270 tensor<int32, [2]> input_21_dilations_0 = const()[name = tensor<string, []>("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1271 tensor<int32, []> input_21_groups_0 = const()[name = tensor<string, []>("input_21_groups_0"), val = tensor<int32, []>(1)];
1272 tensor<fp16, [1536, 384, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10508608)))];
1273 tensor<fp16, [1536]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11688320)))];
1274 tensor<fp16, [1, 1536, 1, 1500]> input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
1275 tensor<string, []> input_23_mode_0 = const()[name = tensor<string, []>("input_23_mode_0"), val = tensor<string, []>("EXACT")];
1276 tensor<fp16, [1, 1536, 1, 1500]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
1277 tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("valid")];
1278 tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = tensor<string, []>("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
1279 tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1280 tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = tensor<string, []>("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1281 tensor<int32, []> hidden_states_9_groups_0 = const()[name = tensor<string, []>("hidden_states_9_groups_0"), val = tensor<int32, []>(1)];
1282 tensor<fp16, [384, 1536, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11691456)))];
1283 tensor<fp16, [384]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12871168)))];
1284 tensor<fp16, [1, 384, 1, 1500]> hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
1285 tensor<fp16, [1, 384, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
1286 tensor<int32, []> var_1717 = const()[name = tensor<string, []>("op_1717"), val = tensor<int32, []>(3)];
1287 tensor<int32, []> var_1728 = const()[name = tensor<string, []>("op_1728"), val = tensor<int32, []>(1)];
1288 tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
1289 tensor<fp16, []> var_1745_to_fp16 = const()[name = tensor<string, []>("op_1745_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1290 tensor<fp16, [1, 384, 1, 1500]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1745_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
1291 tensor<fp16, [384]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12872000)))];
1292 tensor<fp16, [384]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12872832)))];
1293 tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1294 tensor<fp16, [1, 384, 1, 1500]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_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_13_cast_fp16")];
1295 tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("valid")];
1296 tensor<int32, [2]> query_strides_0 = const()[name = tensor<string, []>("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
1297 tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1298 tensor<int32, [2]> query_dilations_0 = const()[name = tensor<string, []>("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1299 tensor<int32, []> query_groups_0 = const()[name = tensor<string, []>("query_groups_0"), val = tensor<int32, []>(1)];
1300 tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12873664)))];
1301 tensor<fp16, [384]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13168640)))];
1302 tensor<fp16, [1, 384, 1, 1500]> query_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
1303 tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")];
1304 tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])];
1305 tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1306 tensor<int32, [2]> key_dilations_0 = const()[name = tensor<string, []>("key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1307 tensor<int32, []> key_groups_0 = const()[name = tensor<string, []>("key_groups_0"), val = tensor<int32, []>(1)];
1308 tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13169472)))];
1309 tensor<fp16, [1, 384, 1, 1500]> key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
1310 tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("valid")];
1311 tensor<int32, [2]> value_strides_0 = const()[name = tensor<string, []>("value_strides_0"), val = tensor<int32, [2]>([1, 1])];
1312 tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1313 tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1314 tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)];
1315 tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13464448)))];
1316 tensor<fp16, [384]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13759424)))];
1317 tensor<fp16, [1, 384, 1, 1500]> value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
1318 tensor<int32, [4]> var_1783_begin_0 = const()[name = tensor<string, []>("op_1783_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1319 tensor<int32, [4]> var_1783_end_0 = const()[name = tensor<string, []>("op_1783_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1320 tensor<bool, [4]> var_1783_end_mask_0 = const()[name = tensor<string, []>("op_1783_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1321 tensor<fp16, [1, 64, 1, 1500]> var_1783_cast_fp16 = slice_by_index(begin = var_1783_begin_0, end = var_1783_end_0, end_mask = var_1783_end_mask_0, x = query_cast_fp16)[name = tensor<string, []>("op_1783_cast_fp16")];
1322 tensor<int32, [4]> var_1787_begin_0 = const()[name = tensor<string, []>("op_1787_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
1323 tensor<int32, [4]> var_1787_end_0 = const()[name = tensor<string, []>("op_1787_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
1324 tensor<bool, [4]> var_1787_end_mask_0 = const()[name = tensor<string, []>("op_1787_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1325 tensor<fp16, [1, 64, 1, 1500]> var_1787_cast_fp16 = slice_by_index(begin = var_1787_begin_0, end = var_1787_end_0, end_mask = var_1787_end_mask_0, x = query_cast_fp16)[name = tensor<string, []>("op_1787_cast_fp16")];
1326 tensor<int32, [4]> var_1791_begin_0 = const()[name = tensor<string, []>("op_1791_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
1327 tensor<int32, [4]> var_1791_end_0 = const()[name = tensor<string, []>("op_1791_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
1328 tensor<bool, [4]> var_1791_end_mask_0 = const()[name = tensor<string, []>("op_1791_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1329 tensor<fp16, [1, 64, 1, 1500]> var_1791_cast_fp16 = slice_by_index(begin = var_1791_begin_0, end = var_1791_end_0, end_mask = var_1791_end_mask_0, x = query_cast_fp16)[name = tensor<string, []>("op_1791_cast_fp16")];
1330 tensor<int32, [4]> var_1795_begin_0 = const()[name = tensor<string, []>("op_1795_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
1331 tensor<int32, [4]> var_1795_end_0 = const()[name = tensor<string, []>("op_1795_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
1332 tensor<bool, [4]> var_1795_end_mask_0 = const()[name = tensor<string, []>("op_1795_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1333 tensor<fp16, [1, 64, 1, 1500]> var_1795_cast_fp16 = slice_by_index(begin = var_1795_begin_0, end = var_1795_end_0, end_mask = var_1795_end_mask_0, x = query_cast_fp16)[name = tensor<string, []>("op_1795_cast_fp16")];
1334 tensor<int32, [4]> var_1799_begin_0 = const()[name = tensor<string, []>("op_1799_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
1335 tensor<int32, [4]> var_1799_end_0 = const()[name = tensor<string, []>("op_1799_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
1336 tensor<bool, [4]> var_1799_end_mask_0 = const()[name = tensor<string, []>("op_1799_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1337 tensor<fp16, [1, 64, 1, 1500]> var_1799_cast_fp16 = slice_by_index(begin = var_1799_begin_0, end = var_1799_end_0, end_mask = var_1799_end_mask_0, x = query_cast_fp16)[name = tensor<string, []>("op_1799_cast_fp16")];
1338 tensor<int32, [4]> var_1803_begin_0 = const()[name = tensor<string, []>("op_1803_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
1339 tensor<int32, [4]> var_1803_end_0 = const()[name = tensor<string, []>("op_1803_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
1340 tensor<bool, [4]> var_1803_end_mask_0 = const()[name = tensor<string, []>("op_1803_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1341 tensor<fp16, [1, 64, 1, 1500]> var_1803_cast_fp16 = slice_by_index(begin = var_1803_begin_0, end = var_1803_end_0, end_mask = var_1803_end_mask_0, x = query_cast_fp16)[name = tensor<string, []>("op_1803_cast_fp16")];
1342 tensor<int32, [4]> var_1812_begin_0 = const()[name = tensor<string, []>("op_1812_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1343 tensor<int32, [4]> var_1812_end_0 = const()[name = tensor<string, []>("op_1812_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1344 tensor<bool, [4]> var_1812_end_mask_0 = const()[name = tensor<string, []>("op_1812_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1345 tensor<fp16, [1, 64, 1, 375]> var_1812_cast_fp16 = slice_by_index(begin = var_1812_begin_0, end = var_1812_end_0, end_mask = var_1812_end_mask_0, x = var_1783_cast_fp16)[name = tensor<string, []>("op_1812_cast_fp16")];
1346 tensor<int32, [4]> var_1819_begin_0 = const()[name = tensor<string, []>("op_1819_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1347 tensor<int32, [4]> var_1819_end_0 = const()[name = tensor<string, []>("op_1819_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1348 tensor<bool, [4]> var_1819_end_mask_0 = const()[name = tensor<string, []>("op_1819_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1349 tensor<fp16, [1, 64, 1, 375]> var_1819_cast_fp16 = slice_by_index(begin = var_1819_begin_0, end = var_1819_end_0, end_mask = var_1819_end_mask_0, x = var_1783_cast_fp16)[name = tensor<string, []>("op_1819_cast_fp16")];
1350 tensor<int32, [4]> var_1826_begin_0 = const()[name = tensor<string, []>("op_1826_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1351 tensor<int32, [4]> var_1826_end_0 = const()[name = tensor<string, []>("op_1826_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1352 tensor<bool, [4]> var_1826_end_mask_0 = const()[name = tensor<string, []>("op_1826_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1353 tensor<fp16, [1, 64, 1, 375]> var_1826_cast_fp16 = slice_by_index(begin = var_1826_begin_0, end = var_1826_end_0, end_mask = var_1826_end_mask_0, x = var_1783_cast_fp16)[name = tensor<string, []>("op_1826_cast_fp16")];
1354 tensor<int32, [4]> var_1833_begin_0 = const()[name = tensor<string, []>("op_1833_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1355 tensor<int32, [4]> var_1833_end_0 = const()[name = tensor<string, []>("op_1833_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1356 tensor<bool, [4]> var_1833_end_mask_0 = const()[name = tensor<string, []>("op_1833_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1357 tensor<fp16, [1, 64, 1, 375]> var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, x = var_1783_cast_fp16)[name = tensor<string, []>("op_1833_cast_fp16")];
1358 tensor<int32, [4]> var_1840_begin_0 = const()[name = tensor<string, []>("op_1840_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1359 tensor<int32, [4]> var_1840_end_0 = const()[name = tensor<string, []>("op_1840_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1360 tensor<bool, [4]> var_1840_end_mask_0 = const()[name = tensor<string, []>("op_1840_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1361 tensor<fp16, [1, 64, 1, 375]> var_1840_cast_fp16 = slice_by_index(begin = var_1840_begin_0, end = var_1840_end_0, end_mask = var_1840_end_mask_0, x = var_1787_cast_fp16)[name = tensor<string, []>("op_1840_cast_fp16")];
1362 tensor<int32, [4]> var_1847_begin_0 = const()[name = tensor<string, []>("op_1847_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1363 tensor<int32, [4]> var_1847_end_0 = const()[name = tensor<string, []>("op_1847_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1364 tensor<bool, [4]> var_1847_end_mask_0 = const()[name = tensor<string, []>("op_1847_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1365 tensor<fp16, [1, 64, 1, 375]> var_1847_cast_fp16 = slice_by_index(begin = var_1847_begin_0, end = var_1847_end_0, end_mask = var_1847_end_mask_0, x = var_1787_cast_fp16)[name = tensor<string, []>("op_1847_cast_fp16")];
1366 tensor<int32, [4]> var_1854_begin_0 = const()[name = tensor<string, []>("op_1854_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1367 tensor<int32, [4]> var_1854_end_0 = const()[name = tensor<string, []>("op_1854_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1368 tensor<bool, [4]> var_1854_end_mask_0 = const()[name = tensor<string, []>("op_1854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1369 tensor<fp16, [1, 64, 1, 375]> var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, x = var_1787_cast_fp16)[name = tensor<string, []>("op_1854_cast_fp16")];
1370 tensor<int32, [4]> var_1861_begin_0 = const()[name = tensor<string, []>("op_1861_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1371 tensor<int32, [4]> var_1861_end_0 = const()[name = tensor<string, []>("op_1861_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1372 tensor<bool, [4]> var_1861_end_mask_0 = const()[name = tensor<string, []>("op_1861_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1373 tensor<fp16, [1, 64, 1, 375]> var_1861_cast_fp16 = slice_by_index(begin = var_1861_begin_0, end = var_1861_end_0, end_mask = var_1861_end_mask_0, x = var_1787_cast_fp16)[name = tensor<string, []>("op_1861_cast_fp16")];
1374 tensor<int32, [4]> var_1868_begin_0 = const()[name = tensor<string, []>("op_1868_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1375 tensor<int32, [4]> var_1868_end_0 = const()[name = tensor<string, []>("op_1868_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1376 tensor<bool, [4]> var_1868_end_mask_0 = const()[name = tensor<string, []>("op_1868_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1377 tensor<fp16, [1, 64, 1, 375]> var_1868_cast_fp16 = slice_by_index(begin = var_1868_begin_0, end = var_1868_end_0, end_mask = var_1868_end_mask_0, x = var_1791_cast_fp16)[name = tensor<string, []>("op_1868_cast_fp16")];
1378 tensor<int32, [4]> var_1875_begin_0 = const()[name = tensor<string, []>("op_1875_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1379 tensor<int32, [4]> var_1875_end_0 = const()[name = tensor<string, []>("op_1875_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1380 tensor<bool, [4]> var_1875_end_mask_0 = const()[name = tensor<string, []>("op_1875_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1381 tensor<fp16, [1, 64, 1, 375]> var_1875_cast_fp16 = slice_by_index(begin = var_1875_begin_0, end = var_1875_end_0, end_mask = var_1875_end_mask_0, x = var_1791_cast_fp16)[name = tensor<string, []>("op_1875_cast_fp16")];
1382 tensor<int32, [4]> var_1882_begin_0 = const()[name = tensor<string, []>("op_1882_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1383 tensor<int32, [4]> var_1882_end_0 = const()[name = tensor<string, []>("op_1882_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1384 tensor<bool, [4]> var_1882_end_mask_0 = const()[name = tensor<string, []>("op_1882_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1385 tensor<fp16, [1, 64, 1, 375]> var_1882_cast_fp16 = slice_by_index(begin = var_1882_begin_0, end = var_1882_end_0, end_mask = var_1882_end_mask_0, x = var_1791_cast_fp16)[name = tensor<string, []>("op_1882_cast_fp16")];
1386 tensor<int32, [4]> var_1889_begin_0 = const()[name = tensor<string, []>("op_1889_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1387 tensor<int32, [4]> var_1889_end_0 = const()[name = tensor<string, []>("op_1889_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1388 tensor<bool, [4]> var_1889_end_mask_0 = const()[name = tensor<string, []>("op_1889_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1389 tensor<fp16, [1, 64, 1, 375]> var_1889_cast_fp16 = slice_by_index(begin = var_1889_begin_0, end = var_1889_end_0, end_mask = var_1889_end_mask_0, x = var_1791_cast_fp16)[name = tensor<string, []>("op_1889_cast_fp16")];
1390 tensor<int32, [4]> var_1896_begin_0 = const()[name = tensor<string, []>("op_1896_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1391 tensor<int32, [4]> var_1896_end_0 = const()[name = tensor<string, []>("op_1896_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1392 tensor<bool, [4]> var_1896_end_mask_0 = const()[name = tensor<string, []>("op_1896_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1393 tensor<fp16, [1, 64, 1, 375]> var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, x = var_1795_cast_fp16)[name = tensor<string, []>("op_1896_cast_fp16")];
1394 tensor<int32, [4]> var_1903_begin_0 = const()[name = tensor<string, []>("op_1903_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1395 tensor<int32, [4]> var_1903_end_0 = const()[name = tensor<string, []>("op_1903_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1396 tensor<bool, [4]> var_1903_end_mask_0 = const()[name = tensor<string, []>("op_1903_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1397 tensor<fp16, [1, 64, 1, 375]> var_1903_cast_fp16 = slice_by_index(begin = var_1903_begin_0, end = var_1903_end_0, end_mask = var_1903_end_mask_0, x = var_1795_cast_fp16)[name = tensor<string, []>("op_1903_cast_fp16")];
1398 tensor<int32, [4]> var_1910_begin_0 = const()[name = tensor<string, []>("op_1910_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1399 tensor<int32, [4]> var_1910_end_0 = const()[name = tensor<string, []>("op_1910_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1400 tensor<bool, [4]> var_1910_end_mask_0 = const()[name = tensor<string, []>("op_1910_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1401 tensor<fp16, [1, 64, 1, 375]> var_1910_cast_fp16 = slice_by_index(begin = var_1910_begin_0, end = var_1910_end_0, end_mask = var_1910_end_mask_0, x = var_1795_cast_fp16)[name = tensor<string, []>("op_1910_cast_fp16")];
1402 tensor<int32, [4]> var_1917_begin_0 = const()[name = tensor<string, []>("op_1917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1403 tensor<int32, [4]> var_1917_end_0 = const()[name = tensor<string, []>("op_1917_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1404 tensor<bool, [4]> var_1917_end_mask_0 = const()[name = tensor<string, []>("op_1917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1405 tensor<fp16, [1, 64, 1, 375]> var_1917_cast_fp16 = slice_by_index(begin = var_1917_begin_0, end = var_1917_end_0, end_mask = var_1917_end_mask_0, x = var_1795_cast_fp16)[name = tensor<string, []>("op_1917_cast_fp16")];
1406 tensor<int32, [4]> var_1924_begin_0 = const()[name = tensor<string, []>("op_1924_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1407 tensor<int32, [4]> var_1924_end_0 = const()[name = tensor<string, []>("op_1924_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1408 tensor<bool, [4]> var_1924_end_mask_0 = const()[name = tensor<string, []>("op_1924_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1409 tensor<fp16, [1, 64, 1, 375]> var_1924_cast_fp16 = slice_by_index(begin = var_1924_begin_0, end = var_1924_end_0, end_mask = var_1924_end_mask_0, x = var_1799_cast_fp16)[name = tensor<string, []>("op_1924_cast_fp16")];
1410 tensor<int32, [4]> var_1931_begin_0 = const()[name = tensor<string, []>("op_1931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1411 tensor<int32, [4]> var_1931_end_0 = const()[name = tensor<string, []>("op_1931_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1412 tensor<bool, [4]> var_1931_end_mask_0 = const()[name = tensor<string, []>("op_1931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1413 tensor<fp16, [1, 64, 1, 375]> var_1931_cast_fp16 = slice_by_index(begin = var_1931_begin_0, end = var_1931_end_0, end_mask = var_1931_end_mask_0, x = var_1799_cast_fp16)[name = tensor<string, []>("op_1931_cast_fp16")];
1414 tensor<int32, [4]> var_1938_begin_0 = const()[name = tensor<string, []>("op_1938_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1415 tensor<int32, [4]> var_1938_end_0 = const()[name = tensor<string, []>("op_1938_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1416 tensor<bool, [4]> var_1938_end_mask_0 = const()[name = tensor<string, []>("op_1938_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1417 tensor<fp16, [1, 64, 1, 375]> var_1938_cast_fp16 = slice_by_index(begin = var_1938_begin_0, end = var_1938_end_0, end_mask = var_1938_end_mask_0, x = var_1799_cast_fp16)[name = tensor<string, []>("op_1938_cast_fp16")];
1418 tensor<int32, [4]> var_1945_begin_0 = const()[name = tensor<string, []>("op_1945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1419 tensor<int32, [4]> var_1945_end_0 = const()[name = tensor<string, []>("op_1945_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1420 tensor<bool, [4]> var_1945_end_mask_0 = const()[name = tensor<string, []>("op_1945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1421 tensor<fp16, [1, 64, 1, 375]> var_1945_cast_fp16 = slice_by_index(begin = var_1945_begin_0, end = var_1945_end_0, end_mask = var_1945_end_mask_0, x = var_1799_cast_fp16)[name = tensor<string, []>("op_1945_cast_fp16")];
1422 tensor<int32, [4]> var_1952_begin_0 = const()[name = tensor<string, []>("op_1952_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1423 tensor<int32, [4]> var_1952_end_0 = const()[name = tensor<string, []>("op_1952_end_0"), val = tensor<int32, [4]>([1, 64, 1, 375])];
1424 tensor<bool, [4]> var_1952_end_mask_0 = const()[name = tensor<string, []>("op_1952_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1425 tensor<fp16, [1, 64, 1, 375]> var_1952_cast_fp16 = slice_by_index(begin = var_1952_begin_0, end = var_1952_end_0, end_mask = var_1952_end_mask_0, x = var_1803_cast_fp16)[name = tensor<string, []>("op_1952_cast_fp16")];
1426 tensor<int32, [4]> var_1959_begin_0 = const()[name = tensor<string, []>("op_1959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 375])];
1427 tensor<int32, [4]> var_1959_end_0 = const()[name = tensor<string, []>("op_1959_end_0"), val = tensor<int32, [4]>([1, 64, 1, 750])];
1428 tensor<bool, [4]> var_1959_end_mask_0 = const()[name = tensor<string, []>("op_1959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1429 tensor<fp16, [1, 64, 1, 375]> var_1959_cast_fp16 = slice_by_index(begin = var_1959_begin_0, end = var_1959_end_0, end_mask = var_1959_end_mask_0, x = var_1803_cast_fp16)[name = tensor<string, []>("op_1959_cast_fp16")];
1430 tensor<int32, [4]> var_1966_begin_0 = const()[name = tensor<string, []>("op_1966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 750])];
1431 tensor<int32, [4]> var_1966_end_0 = const()[name = tensor<string, []>("op_1966_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1125])];
1432 tensor<bool, [4]> var_1966_end_mask_0 = const()[name = tensor<string, []>("op_1966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1433 tensor<fp16, [1, 64, 1, 375]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = var_1803_cast_fp16)[name = tensor<string, []>("op_1966_cast_fp16")];
1434 tensor<int32, [4]> var_1973_begin_0 = const()[name = tensor<string, []>("op_1973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1125])];
1435 tensor<int32, [4]> var_1973_end_0 = const()[name = tensor<string, []>("op_1973_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1436 tensor<bool, [4]> var_1973_end_mask_0 = const()[name = tensor<string, []>("op_1973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1437 tensor<fp16, [1, 64, 1, 375]> var_1973_cast_fp16 = slice_by_index(begin = var_1973_begin_0, end = var_1973_end_0, end_mask = var_1973_end_mask_0, x = var_1803_cast_fp16)[name = tensor<string, []>("op_1973_cast_fp16")];
1438 tensor<int32, [4]> k_perm_0 = const()[name = tensor<string, []>("k_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
1439 tensor<int32, [4]> var_1978_begin_0 = const()[name = tensor<string, []>("op_1978_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1440 tensor<int32, [4]> var_1978_end_0 = const()[name = tensor<string, []>("op_1978_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 64])];
1441 tensor<bool, [4]> var_1978_end_mask_0 = const()[name = tensor<string, []>("op_1978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1442 tensor<fp16, [1, 1500, 1, 384]> k_cast_fp16 = transpose(perm = k_perm_0, x = key_cast_fp16)[name = tensor<string, []>("transpose_0")];
1443 tensor<fp16, [1, 1500, 1, 64]> var_1978_cast_fp16 = slice_by_index(begin = var_1978_begin_0, end = var_1978_end_0, end_mask = var_1978_end_mask_0, x = k_cast_fp16)[name = tensor<string, []>("op_1978_cast_fp16")];
1444 tensor<int32, [4]> var_1982_begin_0 = const()[name = tensor<string, []>("op_1982_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
1445 tensor<int32, [4]> var_1982_end_0 = const()[name = tensor<string, []>("op_1982_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 128])];
1446 tensor<bool, [4]> var_1982_end_mask_0 = const()[name = tensor<string, []>("op_1982_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1447 tensor<fp16, [1, 1500, 1, 64]> var_1982_cast_fp16 = slice_by_index(begin = var_1982_begin_0, end = var_1982_end_0, end_mask = var_1982_end_mask_0, x = k_cast_fp16)[name = tensor<string, []>("op_1982_cast_fp16")];
1448 tensor<int32, [4]> var_1986_begin_0 = const()[name = tensor<string, []>("op_1986_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
1449 tensor<int32, [4]> var_1986_end_0 = const()[name = tensor<string, []>("op_1986_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 192])];
1450 tensor<bool, [4]> var_1986_end_mask_0 = const()[name = tensor<string, []>("op_1986_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1451 tensor<fp16, [1, 1500, 1, 64]> var_1986_cast_fp16 = slice_by_index(begin = var_1986_begin_0, end = var_1986_end_0, end_mask = var_1986_end_mask_0, x = k_cast_fp16)[name = tensor<string, []>("op_1986_cast_fp16")];
1452 tensor<int32, [4]> var_1990_begin_0 = const()[name = tensor<string, []>("op_1990_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 192])];
1453 tensor<int32, [4]> var_1990_end_0 = const()[name = tensor<string, []>("op_1990_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 256])];
1454 tensor<bool, [4]> var_1990_end_mask_0 = const()[name = tensor<string, []>("op_1990_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1455 tensor<fp16, [1, 1500, 1, 64]> var_1990_cast_fp16 = slice_by_index(begin = var_1990_begin_0, end = var_1990_end_0, end_mask = var_1990_end_mask_0, x = k_cast_fp16)[name = tensor<string, []>("op_1990_cast_fp16")];
1456 tensor<int32, [4]> var_1994_begin_0 = const()[name = tensor<string, []>("op_1994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
1457 tensor<int32, [4]> var_1994_end_0 = const()[name = tensor<string, []>("op_1994_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 320])];
1458 tensor<bool, [4]> var_1994_end_mask_0 = const()[name = tensor<string, []>("op_1994_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1459 tensor<fp16, [1, 1500, 1, 64]> var_1994_cast_fp16 = slice_by_index(begin = var_1994_begin_0, end = var_1994_end_0, end_mask = var_1994_end_mask_0, x = k_cast_fp16)[name = tensor<string, []>("op_1994_cast_fp16")];
1460 tensor<int32, [4]> var_1998_begin_0 = const()[name = tensor<string, []>("op_1998_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 320])];
1461 tensor<int32, [4]> var_1998_end_0 = const()[name = tensor<string, []>("op_1998_end_0"), val = tensor<int32, [4]>([1, 1500, 1, 384])];
1462 tensor<bool, [4]> var_1998_end_mask_0 = const()[name = tensor<string, []>("op_1998_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
1463 tensor<fp16, [1, 1500, 1, 64]> var_1998_cast_fp16 = slice_by_index(begin = var_1998_begin_0, end = var_1998_end_0, end_mask = var_1998_end_mask_0, x = k_cast_fp16)[name = tensor<string, []>("op_1998_cast_fp16")];
1464 tensor<int32, [4]> var_2000_begin_0 = const()[name = tensor<string, []>("op_2000_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1465 tensor<int32, [4]> var_2000_end_0 = const()[name = tensor<string, []>("op_2000_end_0"), val = tensor<int32, [4]>([1, 64, 1, 1500])];
1466 tensor<bool, [4]> var_2000_end_mask_0 = const()[name = tensor<string, []>("op_2000_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1467 tensor<fp16, [1, 64, 1, 1500]> var_2000_cast_fp16 = slice_by_index(begin = var_2000_begin_0, end = var_2000_end_0, end_mask = var_2000_end_mask_0, x = value_cast_fp16)[name = tensor<string, []>("op_2000_cast_fp16")];
1468 tensor<int32, [4]> var_2004_begin_0 = const()[name = tensor<string, []>("op_2004_begin_0"), val = tensor<int32, [4]>([0, 64, 0, 0])];
1469 tensor<int32, [4]> var_2004_end_0 = const()[name = tensor<string, []>("op_2004_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1500])];
1470 tensor<bool, [4]> var_2004_end_mask_0 = const()[name = tensor<string, []>("op_2004_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1471 tensor<fp16, [1, 64, 1, 1500]> var_2004_cast_fp16 = slice_by_index(begin = var_2004_begin_0, end = var_2004_end_0, end_mask = var_2004_end_mask_0, x = value_cast_fp16)[name = tensor<string, []>("op_2004_cast_fp16")];
1472 tensor<int32, [4]> var_2008_begin_0 = const()[name = tensor<string, []>("op_2008_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
1473 tensor<int32, [4]> var_2008_end_0 = const()[name = tensor<string, []>("op_2008_end_0"), val = tensor<int32, [4]>([1, 192, 1, 1500])];
1474 tensor<bool, [4]> var_2008_end_mask_0 = const()[name = tensor<string, []>("op_2008_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1475 tensor<fp16, [1, 64, 1, 1500]> var_2008_cast_fp16 = slice_by_index(begin = var_2008_begin_0, end = var_2008_end_0, end_mask = var_2008_end_mask_0, x = value_cast_fp16)[name = tensor<string, []>("op_2008_cast_fp16")];
1476 tensor<int32, [4]> var_2012_begin_0 = const()[name = tensor<string, []>("op_2012_begin_0"), val = tensor<int32, [4]>([0, 192, 0, 0])];
1477 tensor<int32, [4]> var_2012_end_0 = const()[name = tensor<string, []>("op_2012_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1500])];
1478 tensor<bool, [4]> var_2012_end_mask_0 = const()[name = tensor<string, []>("op_2012_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1479 tensor<fp16, [1, 64, 1, 1500]> var_2012_cast_fp16 = slice_by_index(begin = var_2012_begin_0, end = var_2012_end_0, end_mask = var_2012_end_mask_0, x = value_cast_fp16)[name = tensor<string, []>("op_2012_cast_fp16")];
1480 tensor<int32, [4]> var_2016_begin_0 = const()[name = tensor<string, []>("op_2016_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
1481 tensor<int32, [4]> var_2016_end_0 = const()[name = tensor<string, []>("op_2016_end_0"), val = tensor<int32, [4]>([1, 320, 1, 1500])];
1482 tensor<bool, [4]> var_2016_end_mask_0 = const()[name = tensor<string, []>("op_2016_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1483 tensor<fp16, [1, 64, 1, 1500]> var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, x = value_cast_fp16)[name = tensor<string, []>("op_2016_cast_fp16")];
1484 tensor<int32, [4]> var_2020_begin_0 = const()[name = tensor<string, []>("op_2020_begin_0"), val = tensor<int32, [4]>([0, 320, 0, 0])];
1485 tensor<int32, [4]> var_2020_end_0 = const()[name = tensor<string, []>("op_2020_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1500])];
1486 tensor<bool, [4]> var_2020_end_mask_0 = const()[name = tensor<string, []>("op_2020_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
1487 tensor<fp16, [1, 64, 1, 1500]> var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = var_2020_end_0, end_mask = var_2020_end_mask_0, x = value_cast_fp16)[name = tensor<string, []>("op_2020_cast_fp16")];
1488 tensor<string, []> _SplitHeadsQ__mh_w_145_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_145_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1489 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_145_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_145_equation_0, values = (var_1978_cast_fp16, var_1812_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_145_cast_fp16")];
1490 tensor<string, []> _SplitHeadsQ__mh_w_147_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_147_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1491 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_147_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_147_equation_0, values = (var_1978_cast_fp16, var_1819_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_147_cast_fp16")];
1492 tensor<string, []> _SplitHeadsQ__mh_w_149_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_149_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1493 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_149_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_149_equation_0, values = (var_1978_cast_fp16, var_1826_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_149_cast_fp16")];
1494 tensor<string, []> _SplitHeadsQ__mh_w_151_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_151_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1495 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_151_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_151_equation_0, values = (var_1978_cast_fp16, var_1833_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_151_cast_fp16")];
1496 tensor<string, []> _SplitHeadsQ__mh_w_153_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_153_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1497 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_153_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_153_equation_0, values = (var_1982_cast_fp16, var_1840_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_153_cast_fp16")];
1498 tensor<string, []> _SplitHeadsQ__mh_w_155_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_155_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1499 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_155_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_155_equation_0, values = (var_1982_cast_fp16, var_1847_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_155_cast_fp16")];
1500 tensor<string, []> _SplitHeadsQ__mh_w_157_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_157_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1501 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_157_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_157_equation_0, values = (var_1982_cast_fp16, var_1854_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_157_cast_fp16")];
1502 tensor<string, []> _SplitHeadsQ__mh_w_159_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_159_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1503 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_159_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_159_equation_0, values = (var_1982_cast_fp16, var_1861_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_159_cast_fp16")];
1504 tensor<string, []> _SplitHeadsQ__mh_w_161_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_161_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1505 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_161_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_161_equation_0, values = (var_1986_cast_fp16, var_1868_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_161_cast_fp16")];
1506 tensor<string, []> _SplitHeadsQ__mh_w_163_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_163_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1507 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_163_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_163_equation_0, values = (var_1986_cast_fp16, var_1875_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_163_cast_fp16")];
1508 tensor<string, []> _SplitHeadsQ__mh_w_165_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_165_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1509 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_165_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_165_equation_0, values = (var_1986_cast_fp16, var_1882_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_165_cast_fp16")];
1510 tensor<string, []> _SplitHeadsQ__mh_w_167_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_167_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1511 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_167_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_167_equation_0, values = (var_1986_cast_fp16, var_1889_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_167_cast_fp16")];
1512 tensor<string, []> _SplitHeadsQ__mh_w_169_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_169_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1513 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_169_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_169_equation_0, values = (var_1990_cast_fp16, var_1896_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_169_cast_fp16")];
1514 tensor<string, []> _SplitHeadsQ__mh_w_171_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_171_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1515 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_171_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_171_equation_0, values = (var_1990_cast_fp16, var_1903_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_171_cast_fp16")];
1516 tensor<string, []> _SplitHeadsQ__mh_w_173_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_173_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1517 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_173_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_173_equation_0, values = (var_1990_cast_fp16, var_1910_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_173_cast_fp16")];
1518 tensor<string, []> _SplitHeadsQ__mh_w_175_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_175_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1519 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_175_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_175_equation_0, values = (var_1990_cast_fp16, var_1917_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_175_cast_fp16")];
1520 tensor<string, []> _SplitHeadsQ__mh_w_177_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_177_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1521 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_177_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_177_equation_0, values = (var_1994_cast_fp16, var_1924_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_177_cast_fp16")];
1522 tensor<string, []> _SplitHeadsQ__mh_w_179_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_179_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1523 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_179_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_179_equation_0, values = (var_1994_cast_fp16, var_1931_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_179_cast_fp16")];
1524 tensor<string, []> _SplitHeadsQ__mh_w_181_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_181_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1525 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_181_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_181_equation_0, values = (var_1994_cast_fp16, var_1938_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_181_cast_fp16")];
1526 tensor<string, []> _SplitHeadsQ__mh_w_183_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_183_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1527 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_183_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_183_equation_0, values = (var_1994_cast_fp16, var_1945_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_183_cast_fp16")];
1528 tensor<string, []> _SplitHeadsQ__mh_w_185_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_185_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1529 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_185_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_185_equation_0, values = (var_1998_cast_fp16, var_1952_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_185_cast_fp16")];
1530 tensor<string, []> _SplitHeadsQ__mh_w_187_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_187_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1531 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_187_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_187_equation_0, values = (var_1998_cast_fp16, var_1959_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_187_cast_fp16")];
1532 tensor<string, []> _SplitHeadsQ__mh_w_189_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_189_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1533 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_189_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_189_equation_0, values = (var_1998_cast_fp16, var_1966_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_189_cast_fp16")];
1534 tensor<string, []> _SplitHeadsQ__mh_w_equation_0 = const()[name = tensor<string, []>("_SplitHeadsQ__mh_w_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
1535 tensor<fp16, [1, 1500, 1, 375]> _SplitHeadsQ__mh_w_cast_fp16 = einsum(equation = _SplitHeadsQ__mh_w_equation_0, values = (var_1998_cast_fp16, var_1973_cast_fp16))[name = tensor<string, []>("_SplitHeadsQ__mh_w_cast_fp16")];
1536 tensor<fp16, []> var_2071_to_fp16 = const()[name = tensor<string, []>("op_2071_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1537 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_145_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_145_cast_fp16, y = var_2071_to_fp16)[name = tensor<string, []>("aw_chunk_145_cast_fp16")];
1538 tensor<fp16, []> var_2073_to_fp16 = const()[name = tensor<string, []>("op_2073_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1539 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_147_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_147_cast_fp16, y = var_2073_to_fp16)[name = tensor<string, []>("aw_chunk_147_cast_fp16")];
1540 tensor<fp16, []> var_2075_to_fp16 = const()[name = tensor<string, []>("op_2075_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1541 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_149_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_149_cast_fp16, y = var_2075_to_fp16)[name = tensor<string, []>("aw_chunk_149_cast_fp16")];
1542 tensor<fp16, []> var_2077_to_fp16 = const()[name = tensor<string, []>("op_2077_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1543 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_151_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_151_cast_fp16, y = var_2077_to_fp16)[name = tensor<string, []>("aw_chunk_151_cast_fp16")];
1544 tensor<fp16, []> var_2079_to_fp16 = const()[name = tensor<string, []>("op_2079_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1545 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_153_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_153_cast_fp16, y = var_2079_to_fp16)[name = tensor<string, []>("aw_chunk_153_cast_fp16")];
1546 tensor<fp16, []> var_2081_to_fp16 = const()[name = tensor<string, []>("op_2081_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1547 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_155_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_155_cast_fp16, y = var_2081_to_fp16)[name = tensor<string, []>("aw_chunk_155_cast_fp16")];
1548 tensor<fp16, []> var_2083_to_fp16 = const()[name = tensor<string, []>("op_2083_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1549 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_157_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_157_cast_fp16, y = var_2083_to_fp16)[name = tensor<string, []>("aw_chunk_157_cast_fp16")];
1550 tensor<fp16, []> var_2085_to_fp16 = const()[name = tensor<string, []>("op_2085_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1551 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_159_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_159_cast_fp16, y = var_2085_to_fp16)[name = tensor<string, []>("aw_chunk_159_cast_fp16")];
1552 tensor<fp16, []> var_2087_to_fp16 = const()[name = tensor<string, []>("op_2087_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1553 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_161_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_161_cast_fp16, y = var_2087_to_fp16)[name = tensor<string, []>("aw_chunk_161_cast_fp16")];
1554 tensor<fp16, []> var_2089_to_fp16 = const()[name = tensor<string, []>("op_2089_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1555 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_163_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_163_cast_fp16, y = var_2089_to_fp16)[name = tensor<string, []>("aw_chunk_163_cast_fp16")];
1556 tensor<fp16, []> var_2091_to_fp16 = const()[name = tensor<string, []>("op_2091_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1557 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_165_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_165_cast_fp16, y = var_2091_to_fp16)[name = tensor<string, []>("aw_chunk_165_cast_fp16")];
1558 tensor<fp16, []> var_2093_to_fp16 = const()[name = tensor<string, []>("op_2093_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1559 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_167_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_167_cast_fp16, y = var_2093_to_fp16)[name = tensor<string, []>("aw_chunk_167_cast_fp16")];
1560 tensor<fp16, []> var_2095_to_fp16 = const()[name = tensor<string, []>("op_2095_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1561 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_169_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_169_cast_fp16, y = var_2095_to_fp16)[name = tensor<string, []>("aw_chunk_169_cast_fp16")];
1562 tensor<fp16, []> var_2097_to_fp16 = const()[name = tensor<string, []>("op_2097_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1563 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_171_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_171_cast_fp16, y = var_2097_to_fp16)[name = tensor<string, []>("aw_chunk_171_cast_fp16")];
1564 tensor<fp16, []> var_2099_to_fp16 = const()[name = tensor<string, []>("op_2099_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1565 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_173_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_173_cast_fp16, y = var_2099_to_fp16)[name = tensor<string, []>("aw_chunk_173_cast_fp16")];
1566 tensor<fp16, []> var_2101_to_fp16 = const()[name = tensor<string, []>("op_2101_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1567 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_175_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_175_cast_fp16, y = var_2101_to_fp16)[name = tensor<string, []>("aw_chunk_175_cast_fp16")];
1568 tensor<fp16, []> var_2103_to_fp16 = const()[name = tensor<string, []>("op_2103_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1569 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_177_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_177_cast_fp16, y = var_2103_to_fp16)[name = tensor<string, []>("aw_chunk_177_cast_fp16")];
1570 tensor<fp16, []> var_2105_to_fp16 = const()[name = tensor<string, []>("op_2105_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1571 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_179_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_179_cast_fp16, y = var_2105_to_fp16)[name = tensor<string, []>("aw_chunk_179_cast_fp16")];
1572 tensor<fp16, []> var_2107_to_fp16 = const()[name = tensor<string, []>("op_2107_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1573 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_181_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_181_cast_fp16, y = var_2107_to_fp16)[name = tensor<string, []>("aw_chunk_181_cast_fp16")];
1574 tensor<fp16, []> var_2109_to_fp16 = const()[name = tensor<string, []>("op_2109_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1575 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_183_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_183_cast_fp16, y = var_2109_to_fp16)[name = tensor<string, []>("aw_chunk_183_cast_fp16")];
1576 tensor<fp16, []> var_2111_to_fp16 = const()[name = tensor<string, []>("op_2111_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1577 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_185_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_185_cast_fp16, y = var_2111_to_fp16)[name = tensor<string, []>("aw_chunk_185_cast_fp16")];
1578 tensor<fp16, []> var_2113_to_fp16 = const()[name = tensor<string, []>("op_2113_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1579 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_187_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_187_cast_fp16, y = var_2113_to_fp16)[name = tensor<string, []>("aw_chunk_187_cast_fp16")];
1580 tensor<fp16, []> var_2115_to_fp16 = const()[name = tensor<string, []>("op_2115_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1581 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_189_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_189_cast_fp16, y = var_2115_to_fp16)[name = tensor<string, []>("aw_chunk_189_cast_fp16")];
1582 tensor<fp16, []> var_2117_to_fp16 = const()[name = tensor<string, []>("op_2117_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
1583 tensor<fp16, [1, 1500, 1, 375]> aw_chunk_cast_fp16 = mul(x = _SplitHeadsQ__mh_w_cast_fp16, y = var_2117_to_fp16)[name = tensor<string, []>("aw_chunk_cast_fp16")];
1584 tensor<fp16, [1, 1500, 1, 375]> var_2119_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_145_cast_fp16)[name = tensor<string, []>("op_2119_cast_fp16")];
1585 tensor<fp16, [1, 1500, 1, 375]> var_2120_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_147_cast_fp16)[name = tensor<string, []>("op_2120_cast_fp16")];
1586 tensor<fp16, [1, 1500, 1, 375]> var_2121_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_149_cast_fp16)[name = tensor<string, []>("op_2121_cast_fp16")];
1587 tensor<fp16, [1, 1500, 1, 375]> var_2122_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_151_cast_fp16)[name = tensor<string, []>("op_2122_cast_fp16")];
1588 tensor<fp16, [1, 1500, 1, 375]> var_2123_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_153_cast_fp16)[name = tensor<string, []>("op_2123_cast_fp16")];
1589 tensor<fp16, [1, 1500, 1, 375]> var_2124_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_155_cast_fp16)[name = tensor<string, []>("op_2124_cast_fp16")];
1590 tensor<fp16, [1, 1500, 1, 375]> var_2125_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_157_cast_fp16)[name = tensor<string, []>("op_2125_cast_fp16")];
1591 tensor<fp16, [1, 1500, 1, 375]> var_2126_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_159_cast_fp16)[name = tensor<string, []>("op_2126_cast_fp16")];
1592 tensor<fp16, [1, 1500, 1, 375]> var_2127_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_161_cast_fp16)[name = tensor<string, []>("op_2127_cast_fp16")];
1593 tensor<fp16, [1, 1500, 1, 375]> var_2128_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_163_cast_fp16)[name = tensor<string, []>("op_2128_cast_fp16")];
1594 tensor<fp16, [1, 1500, 1, 375]> var_2129_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_165_cast_fp16)[name = tensor<string, []>("op_2129_cast_fp16")];
1595 tensor<fp16, [1, 1500, 1, 375]> var_2130_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_167_cast_fp16)[name = tensor<string, []>("op_2130_cast_fp16")];
1596 tensor<fp16, [1, 1500, 1, 375]> var_2131_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_169_cast_fp16)[name = tensor<string, []>("op_2131_cast_fp16")];
1597 tensor<fp16, [1, 1500, 1, 375]> var_2132_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_171_cast_fp16)[name = tensor<string, []>("op_2132_cast_fp16")];
1598 tensor<fp16, [1, 1500, 1, 375]> var_2133_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_173_cast_fp16)[name = tensor<string, []>("op_2133_cast_fp16")];
1599 tensor<fp16, [1, 1500, 1, 375]> var_2134_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_175_cast_fp16)[name = tensor<string, []>("op_2134_cast_fp16")];
1600 tensor<fp16, [1, 1500, 1, 375]> var_2135_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_177_cast_fp16)[name = tensor<string, []>("op_2135_cast_fp16")];
1601 tensor<fp16, [1, 1500, 1, 375]> var_2136_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_179_cast_fp16)[name = tensor<string, []>("op_2136_cast_fp16")];
1602 tensor<fp16, [1, 1500, 1, 375]> var_2137_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_181_cast_fp16)[name = tensor<string, []>("op_2137_cast_fp16")];
1603 tensor<fp16, [1, 1500, 1, 375]> var_2138_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_183_cast_fp16)[name = tensor<string, []>("op_2138_cast_fp16")];
1604 tensor<fp16, [1, 1500, 1, 375]> var_2139_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_185_cast_fp16)[name = tensor<string, []>("op_2139_cast_fp16")];
1605 tensor<fp16, [1, 1500, 1, 375]> var_2140_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_187_cast_fp16)[name = tensor<string, []>("op_2140_cast_fp16")];
1606 tensor<fp16, [1, 1500, 1, 375]> var_2141_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_189_cast_fp16)[name = tensor<string, []>("op_2141_cast_fp16")];
1607 tensor<fp16, [1, 1500, 1, 375]> var_2142_cast_fp16 = softmax(axis = var_1728, x = aw_chunk_cast_fp16)[name = tensor<string, []>("op_2142_cast_fp16")];
1608 tensor<string, []> var_2144_equation_0 = const()[name = tensor<string, []>("op_2144_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1609 tensor<fp16, [1, 64, 1, 375]> var_2144_cast_fp16 = einsum(equation = var_2144_equation_0, values = (var_2000_cast_fp16, var_2119_cast_fp16))[name = tensor<string, []>("op_2144_cast_fp16")];
1610 tensor<string, []> var_2146_equation_0 = const()[name = tensor<string, []>("op_2146_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1611 tensor<fp16, [1, 64, 1, 375]> var_2146_cast_fp16 = einsum(equation = var_2146_equation_0, values = (var_2000_cast_fp16, var_2120_cast_fp16))[name = tensor<string, []>("op_2146_cast_fp16")];
1612 tensor<string, []> var_2148_equation_0 = const()[name = tensor<string, []>("op_2148_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1613 tensor<fp16, [1, 64, 1, 375]> var_2148_cast_fp16 = einsum(equation = var_2148_equation_0, values = (var_2000_cast_fp16, var_2121_cast_fp16))[name = tensor<string, []>("op_2148_cast_fp16")];
1614 tensor<string, []> var_2150_equation_0 = const()[name = tensor<string, []>("op_2150_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1615 tensor<fp16, [1, 64, 1, 375]> var_2150_cast_fp16 = einsum(equation = var_2150_equation_0, values = (var_2000_cast_fp16, var_2122_cast_fp16))[name = tensor<string, []>("op_2150_cast_fp16")];
1616 tensor<string, []> var_2152_equation_0 = const()[name = tensor<string, []>("op_2152_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1617 tensor<fp16, [1, 64, 1, 375]> var_2152_cast_fp16 = einsum(equation = var_2152_equation_0, values = (var_2004_cast_fp16, var_2123_cast_fp16))[name = tensor<string, []>("op_2152_cast_fp16")];
1618 tensor<string, []> var_2154_equation_0 = const()[name = tensor<string, []>("op_2154_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1619 tensor<fp16, [1, 64, 1, 375]> var_2154_cast_fp16 = einsum(equation = var_2154_equation_0, values = (var_2004_cast_fp16, var_2124_cast_fp16))[name = tensor<string, []>("op_2154_cast_fp16")];
1620 tensor<string, []> var_2156_equation_0 = const()[name = tensor<string, []>("op_2156_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1621 tensor<fp16, [1, 64, 1, 375]> var_2156_cast_fp16 = einsum(equation = var_2156_equation_0, values = (var_2004_cast_fp16, var_2125_cast_fp16))[name = tensor<string, []>("op_2156_cast_fp16")];
1622 tensor<string, []> var_2158_equation_0 = const()[name = tensor<string, []>("op_2158_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1623 tensor<fp16, [1, 64, 1, 375]> var_2158_cast_fp16 = einsum(equation = var_2158_equation_0, values = (var_2004_cast_fp16, var_2126_cast_fp16))[name = tensor<string, []>("op_2158_cast_fp16")];
1624 tensor<string, []> var_2160_equation_0 = const()[name = tensor<string, []>("op_2160_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1625 tensor<fp16, [1, 64, 1, 375]> var_2160_cast_fp16 = einsum(equation = var_2160_equation_0, values = (var_2008_cast_fp16, var_2127_cast_fp16))[name = tensor<string, []>("op_2160_cast_fp16")];
1626 tensor<string, []> var_2162_equation_0 = const()[name = tensor<string, []>("op_2162_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1627 tensor<fp16, [1, 64, 1, 375]> var_2162_cast_fp16 = einsum(equation = var_2162_equation_0, values = (var_2008_cast_fp16, var_2128_cast_fp16))[name = tensor<string, []>("op_2162_cast_fp16")];
1628 tensor<string, []> var_2164_equation_0 = const()[name = tensor<string, []>("op_2164_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1629 tensor<fp16, [1, 64, 1, 375]> var_2164_cast_fp16 = einsum(equation = var_2164_equation_0, values = (var_2008_cast_fp16, var_2129_cast_fp16))[name = tensor<string, []>("op_2164_cast_fp16")];
1630 tensor<string, []> var_2166_equation_0 = const()[name = tensor<string, []>("op_2166_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1631 tensor<fp16, [1, 64, 1, 375]> var_2166_cast_fp16 = einsum(equation = var_2166_equation_0, values = (var_2008_cast_fp16, var_2130_cast_fp16))[name = tensor<string, []>("op_2166_cast_fp16")];
1632 tensor<string, []> var_2168_equation_0 = const()[name = tensor<string, []>("op_2168_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1633 tensor<fp16, [1, 64, 1, 375]> var_2168_cast_fp16 = einsum(equation = var_2168_equation_0, values = (var_2012_cast_fp16, var_2131_cast_fp16))[name = tensor<string, []>("op_2168_cast_fp16")];
1634 tensor<string, []> var_2170_equation_0 = const()[name = tensor<string, []>("op_2170_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1635 tensor<fp16, [1, 64, 1, 375]> var_2170_cast_fp16 = einsum(equation = var_2170_equation_0, values = (var_2012_cast_fp16, var_2132_cast_fp16))[name = tensor<string, []>("op_2170_cast_fp16")];
1636 tensor<string, []> var_2172_equation_0 = const()[name = tensor<string, []>("op_2172_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1637 tensor<fp16, [1, 64, 1, 375]> var_2172_cast_fp16 = einsum(equation = var_2172_equation_0, values = (var_2012_cast_fp16, var_2133_cast_fp16))[name = tensor<string, []>("op_2172_cast_fp16")];
1638 tensor<string, []> var_2174_equation_0 = const()[name = tensor<string, []>("op_2174_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1639 tensor<fp16, [1, 64, 1, 375]> var_2174_cast_fp16 = einsum(equation = var_2174_equation_0, values = (var_2012_cast_fp16, var_2134_cast_fp16))[name = tensor<string, []>("op_2174_cast_fp16")];
1640 tensor<string, []> var_2176_equation_0 = const()[name = tensor<string, []>("op_2176_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1641 tensor<fp16, [1, 64, 1, 375]> var_2176_cast_fp16 = einsum(equation = var_2176_equation_0, values = (var_2016_cast_fp16, var_2135_cast_fp16))[name = tensor<string, []>("op_2176_cast_fp16")];
1642 tensor<string, []> var_2178_equation_0 = const()[name = tensor<string, []>("op_2178_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1643 tensor<fp16, [1, 64, 1, 375]> var_2178_cast_fp16 = einsum(equation = var_2178_equation_0, values = (var_2016_cast_fp16, var_2136_cast_fp16))[name = tensor<string, []>("op_2178_cast_fp16")];
1644 tensor<string, []> var_2180_equation_0 = const()[name = tensor<string, []>("op_2180_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1645 tensor<fp16, [1, 64, 1, 375]> var_2180_cast_fp16 = einsum(equation = var_2180_equation_0, values = (var_2016_cast_fp16, var_2137_cast_fp16))[name = tensor<string, []>("op_2180_cast_fp16")];
1646 tensor<string, []> var_2182_equation_0 = const()[name = tensor<string, []>("op_2182_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1647 tensor<fp16, [1, 64, 1, 375]> var_2182_cast_fp16 = einsum(equation = var_2182_equation_0, values = (var_2016_cast_fp16, var_2138_cast_fp16))[name = tensor<string, []>("op_2182_cast_fp16")];
1648 tensor<string, []> var_2184_equation_0 = const()[name = tensor<string, []>("op_2184_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1649 tensor<fp16, [1, 64, 1, 375]> var_2184_cast_fp16 = einsum(equation = var_2184_equation_0, values = (var_2020_cast_fp16, var_2139_cast_fp16))[name = tensor<string, []>("op_2184_cast_fp16")];
1650 tensor<string, []> var_2186_equation_0 = const()[name = tensor<string, []>("op_2186_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1651 tensor<fp16, [1, 64, 1, 375]> var_2186_cast_fp16 = einsum(equation = var_2186_equation_0, values = (var_2020_cast_fp16, var_2140_cast_fp16))[name = tensor<string, []>("op_2186_cast_fp16")];
1652 tensor<string, []> var_2188_equation_0 = const()[name = tensor<string, []>("op_2188_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1653 tensor<fp16, [1, 64, 1, 375]> var_2188_cast_fp16 = einsum(equation = var_2188_equation_0, values = (var_2020_cast_fp16, var_2141_cast_fp16))[name = tensor<string, []>("op_2188_cast_fp16")];
1654 tensor<string, []> var_2190_equation_0 = const()[name = tensor<string, []>("op_2190_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
1655 tensor<fp16, [1, 64, 1, 375]> var_2190_cast_fp16 = einsum(equation = var_2190_equation_0, values = (var_2020_cast_fp16, var_2142_cast_fp16))[name = tensor<string, []>("op_2190_cast_fp16")];
1656 tensor<bool, []> var_2192_interleave_0 = const()[name = tensor<string, []>("op_2192_interleave_0"), val = tensor<bool, []>(false)];
1657 tensor<fp16, [1, 64, 1, 1500]> var_2192_cast_fp16 = concat(axis = var_1717, interleave = var_2192_interleave_0, values = (var_2144_cast_fp16, var_2146_cast_fp16, var_2148_cast_fp16, var_2150_cast_fp16))[name = tensor<string, []>("op_2192_cast_fp16")];
1658 tensor<bool, []> var_2194_interleave_0 = const()[name = tensor<string, []>("op_2194_interleave_0"), val = tensor<bool, []>(false)];
1659 tensor<fp16, [1, 64, 1, 1500]> var_2194_cast_fp16 = concat(axis = var_1717, interleave = var_2194_interleave_0, values = (var_2152_cast_fp16, var_2154_cast_fp16, var_2156_cast_fp16, var_2158_cast_fp16))[name = tensor<string, []>("op_2194_cast_fp16")];
1660 tensor<bool, []> var_2196_interleave_0 = const()[name = tensor<string, []>("op_2196_interleave_0"), val = tensor<bool, []>(false)];
1661 tensor<fp16, [1, 64, 1, 1500]> var_2196_cast_fp16 = concat(axis = var_1717, interleave = var_2196_interleave_0, values = (var_2160_cast_fp16, var_2162_cast_fp16, var_2164_cast_fp16, var_2166_cast_fp16))[name = tensor<string, []>("op_2196_cast_fp16")];
1662 tensor<bool, []> var_2198_interleave_0 = const()[name = tensor<string, []>("op_2198_interleave_0"), val = tensor<bool, []>(false)];
1663 tensor<fp16, [1, 64, 1, 1500]> var_2198_cast_fp16 = concat(axis = var_1717, interleave = var_2198_interleave_0, values = (var_2168_cast_fp16, var_2170_cast_fp16, var_2172_cast_fp16, var_2174_cast_fp16))[name = tensor<string, []>("op_2198_cast_fp16")];
1664 tensor<bool, []> var_2200_interleave_0 = const()[name = tensor<string, []>("op_2200_interleave_0"), val = tensor<bool, []>(false)];
1665 tensor<fp16, [1, 64, 1, 1500]> var_2200_cast_fp16 = concat(axis = var_1717, interleave = var_2200_interleave_0, values = (var_2176_cast_fp16, var_2178_cast_fp16, var_2180_cast_fp16, var_2182_cast_fp16))[name = tensor<string, []>("op_2200_cast_fp16")];
1666 tensor<bool, []> var_2202_interleave_0 = const()[name = tensor<string, []>("op_2202_interleave_0"), val = tensor<bool, []>(false)];
1667 tensor<fp16, [1, 64, 1, 1500]> var_2202_cast_fp16 = concat(axis = var_1717, interleave = var_2202_interleave_0, values = (var_2184_cast_fp16, var_2186_cast_fp16, var_2188_cast_fp16, var_2190_cast_fp16))[name = tensor<string, []>("op_2202_cast_fp16")];
1668 tensor<bool, []> input_25_interleave_0 = const()[name = tensor<string, []>("input_25_interleave_0"), val = tensor<bool, []>(false)];
1669 tensor<fp16, [1, 384, 1, 1500]> input_25_cast_fp16 = concat(axis = var_1728, interleave = input_25_interleave_0, values = (var_2192_cast_fp16, var_2194_cast_fp16, var_2196_cast_fp16, var_2198_cast_fp16, var_2200_cast_fp16, var_2202_cast_fp16))[name = tensor<string, []>("input_25_cast_fp16")];
1670 tensor<string, []> obj_pad_type_0 = const()[name = tensor<string, []>("obj_pad_type_0"), val = tensor<string, []>("valid")];
1671 tensor<int32, [2]> obj_strides_0 = const()[name = tensor<string, []>("obj_strides_0"), val = tensor<int32, [2]>([1, 1])];
1672 tensor<int32, [4]> obj_pad_0 = const()[name = tensor<string, []>("obj_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1673 tensor<int32, [2]> obj_dilations_0 = const()[name = tensor<string, []>("obj_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1674 tensor<int32, []> obj_groups_0 = const()[name = tensor<string, []>("obj_groups_0"), val = tensor<int32, []>(1)];
1675 tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13760256)))];
1676 tensor<fp16, [384]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14055232)))];
1677 tensor<fp16, [1, 384, 1, 1500]> obj_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_dilations_0, groups = obj_groups_0, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = obj_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
1678 tensor<fp16, [1, 384, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
1679 tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
1680 tensor<fp16, []> var_2221_to_fp16 = const()[name = tensor<string, []>("op_2221_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1681 tensor<fp16, [1, 384, 1, 1500]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_2221_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
1682 tensor<fp16, [384]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14056064)))];
1683 tensor<fp16, [384]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14056896)))];
1684 tensor<fp16, []> input_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1685 tensor<fp16, [1, 384, 1, 1500]> input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_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, []>("input_27_cast_fp16")];
1686 tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")];
1687 tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
1688 tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1689 tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1690 tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
1691 tensor<fp16, [1536, 384, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14057728)))];
1692 tensor<fp16, [1536]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15237440)))];
1693 tensor<fp16, [1, 1536, 1, 1500]> input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, 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_3_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
1694 tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
1695 tensor<fp16, [1, 1536, 1, 1500]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
1696 tensor<string, []> hidden_states_pad_type_0 = const()[name = tensor<string, []>("hidden_states_pad_type_0"), val = tensor<string, []>("valid")];
1697 tensor<int32, [2]> hidden_states_strides_0 = const()[name = tensor<string, []>("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
1698 tensor<int32, [4]> hidden_states_pad_0 = const()[name = tensor<string, []>("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
1699 tensor<int32, [2]> hidden_states_dilations_0 = const()[name = tensor<string, []>("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
1700 tensor<int32, []> hidden_states_groups_0 = const()[name = tensor<string, []>("hidden_states_groups_0"), val = tensor<int32, []>(1)];
1701 tensor<fp16, [384, 1536, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15240576)))];
1702 tensor<fp16, [384]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16420288)))];
1703 tensor<fp16, [1, 384, 1, 1500]> hidden_states_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
1704 tensor<fp16, [1, 384, 1, 1500]> inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
1705 tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
1706 tensor<fp16, []> var_2259_to_fp16 = const()[name = tensor<string, []>("op_2259_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1707 tensor<fp16, [1, 384, 1, 1500]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_2259_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
1708 tensor<fp16, [384]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16421120)))];
1709 tensor<fp16, [384]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16421952)))];
1710 tensor<fp16, []> encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
1711 tensor<fp16, [1, 384, 1, 1500]> encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_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, []>("encoder_output_embeds_type_fp32_cast_fp16")];
1712 } -> (encoder_output_embeds);
1713 }