program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<fp16, [1, 110, 768]> audio_embed, tensor<fp32, [1, 256]> encoder_mask, tensor<fp16, [1, 256, 768]> encoder_output) {
            tensor<string, []> cast_94_dtype_0 = const()[name = tensor<string, []>("cast_94_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, 256]> cast_94 = cast(dtype = cast_94_dtype_0, x = encoder_mask)[name = tensor<string, []>("cast_94")];
            tensor<fp16, [1, 110, 768]> var_40_to_fp16 = const()[name = tensor<string, []>("op_40_to_fp16"), val = tensor<fp16, [1, 110, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1, 110, 768]> input_3_cast_fp16 = add(x = audio_embed, y = var_40_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<int32, []> var_54 = const()[name = tensor<string, []>("op_54"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_0_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169088)))];
            tensor<fp16, []> var_61_to_fp16 = const()[name = tensor<string, []>("op_61_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, epsilon = var_61_to_fp16, gamma = layers_0_norm_sa_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_0_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170688)))];
            tensor<fp16, [2304]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3709696)))];
            tensor<fp16, [1, 110, 2304]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_qkv_proj_weight_to_fp16, x = x_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, [5]> var_79 = const()[name = tensor<string, []>("op_79"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_1_cast_fp16 = reshape(shape = var_79, x = linear_0_cast_fp16)[name = tensor<string, []>("qkv_1_cast_fp16")];
            tensor<int32, [5]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_1_squeeze_mask_0 = const()[name = tensor<string, []>("q_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
            tensor<int32, [5]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
            tensor<int32, [5]> v_1_begin_0 = const()[name = tensor<string, []>("v_1_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_1_end_0 = const()[name = tensor<string, []>("v_1_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_1_end_mask_0 = const()[name = tensor<string, []>("v_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
            tensor<int32, [4]> v_t_1_perm_0 = const()[name = tensor<string, []>("v_t_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_94_transpose_x_0 = const()[name = tensor<string, []>("op_94_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_94_transpose_y_0 = const()[name = tensor<string, []>("op_94_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_67 = transpose(perm = transpose_67_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_217")];
            tensor<fp16, [1, 12, 110, 64]> transpose_66 = transpose(perm = transpose_66_perm_0, x = q_1_cast_fp16)[name = tensor<string, []>("transpose_218")];
            tensor<fp16, [1, 12, 110, 110]> var_94_cast_fp16 = matmul(transpose_x = var_94_transpose_x_0, transpose_y = var_94_transpose_y_0, x = transpose_66, y = transpose_67)[name = tensor<string, []>("op_94_cast_fp16")];
            tensor<fp16, []> var_95_to_fp16 = const()[name = tensor<string, []>("op_95_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_1_cast_fp16 = mul(x = var_94_cast_fp16, y = var_95_to_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<string, []> attn_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [110, 110]> var_102 = const()[name = tensor<string, []>("op_102"), val = tensor<fp32, [110, 110]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3714368)))];
            tensor<fp32, [1, 12, 110, 110]> attn_1_cast_fp16_to_fp32 = cast(dtype = attn_1_cast_fp16_to_fp32_dtype_0, x = attn_1_cast_fp16)[name = tensor<string, []>("cast_93")];
            tensor<fp32, [1, 12, 110, 110]> input_5 = add(x = attn_1_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_5")];
            tensor<string, []> input_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_5_to_fp16 = cast(dtype = input_5_to_fp16_dtype_0, x = input_5)[name = tensor<string, []>("cast_92")];
            tensor<fp16, [1, 12, 110, 110]> attn_3_cast_fp16 = softmax(axis = var_54, x = input_5_to_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
            tensor<bool, []> out_1_transpose_x_0 = const()[name = tensor<string, []>("out_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_1_transpose_y_0 = const()[name = tensor<string, []>("out_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_1_cast_fp16 = transpose(perm = v_t_1_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_219")];
            tensor<fp16, [1, 12, 110, 64]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_3_cast_fp16, y = v_t_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<int32, [4]> var_106_perm_0 = const()[name = tensor<string, []>("op_106_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_107 = const()[name = tensor<string, []>("op_107"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_106_cast_fp16 = transpose(perm = var_106_perm_0, x = out_1_cast_fp16)[name = tensor<string, []>("transpose_216")];
            tensor<fp16, [1, 110, 768]> input_7_cast_fp16 = reshape(shape = var_107, x = var_106_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<fp16, [768, 768]> 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, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3762880)))];
            tensor<fp16, [768]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4942592)))];
            tensor<fp16, [1, 110, 768]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_9_cast_fp16 = add(x = input_3_cast_fp16, y = linear_1_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<int32, [1]> x_3_axes_0 = const()[name = tensor<string, []>("x_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_0_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4944192)))];
            tensor<fp16, [1, 110, 768]> x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, epsilon = var_61_to_fp16, gamma = layers_0_norm_xa_query_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<int32, [1]> memory_1_axes_0 = const()[name = tensor<string, []>("memory_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_0_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4945792)))];
            tensor<fp16, [1, 256, 768]> memory_1_cast_fp16 = layer_norm(axes = memory_1_axes_0, epsilon = var_61_to_fp16, gamma = layers_0_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_1_cast_fp16")];
            tensor<fp16, [128, 768]> layers_0_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4947392)))];
            tensor<fp16, [128]> linear_2_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_2_bias_0_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5144064)))];
            tensor<fp16, [1, 110, 128]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_cross_attn_q_proj_weight_to_fp16, x = x_3_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<int32, [4]> var_130 = const()[name = tensor<string, []>("op_130"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_131_cast_fp16 = reshape(shape = var_130, x = linear_2_cast_fp16)[name = tensor<string, []>("op_131_cast_fp16")];
            tensor<fp16, [256, 768]> layers_0_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5144384)))];
            tensor<fp16, [256]> linear_3_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_3_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5537664)))];
            tensor<fp16, [1, 256, 256]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_0_cross_attn_kv_proj_weight_to_fp16, x = memory_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<int32, [5]> var_135 = const()[name = tensor<string, []>("op_135"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_1_cast_fp16 = reshape(shape = var_135, x = linear_3_cast_fp16)[name = tensor<string, []>("kv_1_cast_fp16")];
            tensor<int32, [5]> var_139_begin_0 = const()[name = tensor<string, []>("op_139_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_139_end_0 = const()[name = tensor<string, []>("op_139_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_139_end_mask_0 = const()[name = tensor<string, []>("op_139_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_139_squeeze_mask_0 = const()[name = tensor<string, []>("op_139_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_139_cast_fp16 = slice_by_index(begin = var_139_begin_0, end = var_139_end_0, end_mask = var_139_end_mask_0, squeeze_mask = var_139_squeeze_mask_0, x = kv_1_cast_fp16)[name = tensor<string, []>("op_139_cast_fp16")];
            tensor<int32, [5]> var_143_begin_0 = const()[name = tensor<string, []>("op_143_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_143_end_0 = const()[name = tensor<string, []>("op_143_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_143_end_mask_0 = const()[name = tensor<string, []>("op_143_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_143_squeeze_mask_0 = const()[name = tensor<string, []>("op_143_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_143_cast_fp16 = slice_by_index(begin = var_143_begin_0, end = var_143_end_0, end_mask = var_143_end_mask_0, squeeze_mask = var_143_squeeze_mask_0, x = kv_1_cast_fp16)[name = tensor<string, []>("op_143_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_146_transpose_x_0 = const()[name = tensor<string, []>("op_146_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_146_transpose_y_0 = const()[name = tensor<string, []>("op_146_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_139_cast_fp16)[name = tensor<string, []>("transpose_213")];
            tensor<fp16, [1, 1, 110, 128]> transpose_68 = transpose(perm = transpose_68_perm_0, x = var_131_cast_fp16)[name = tensor<string, []>("transpose_214")];
            tensor<fp16, [1, 1, 110, 256]> var_146_cast_fp16 = matmul(transpose_x = var_146_transpose_x_0, transpose_y = var_146_transpose_y_0, x = transpose_68, y = transpose_69)[name = tensor<string, []>("op_146_cast_fp16")];
            tensor<fp16, []> var_147_to_fp16 = const()[name = tensor<string, []>("op_147_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_5_cast_fp16 = mul(x = var_146_cast_fp16, y = var_147_to_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<int32, [1]> var_149_axes_0 = const()[name = tensor<string, []>("op_149_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 256]> var_149 = expand_dims(axes = var_149_axes_0, x = cast_94)[name = tensor<string, []>("op_149")];
            tensor<int32, [1]> attn_mask_1_axes_0 = const()[name = tensor<string, []>("attn_mask_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, [1, 1, 1, 256]> attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_149)[name = tensor<string, []>("attn_mask_1")];
            tensor<bool, [1, 1, 1, 256]> var_151 = logical_not(x = attn_mask_1)[name = tensor<string, []>("op_151")];
            tensor<fp16, []> var_45_to_fp16 = const()[name = tensor<string, []>("op_45_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_11_cast_fp16 = select(a = var_45_to_fp16, b = attn_5_cast_fp16, cond = var_151)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_7_cast_fp16 = softmax(axis = var_54, x = input_11_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
            tensor<bool, []> out_3_transpose_x_0 = const()[name = tensor<string, []>("out_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_3_transpose_y_0 = const()[name = tensor<string, []>("out_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_143_cast_fp16)[name = tensor<string, []>("transpose_215")];
            tensor<fp16, [1, 1, 110, 128]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_7_cast_fp16, y = v_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<int32, [4]> var_155_perm_0 = const()[name = tensor<string, []>("op_155_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_156 = const()[name = tensor<string, []>("op_156"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_155_cast_fp16 = transpose(perm = var_155_perm_0, x = out_3_cast_fp16)[name = tensor<string, []>("transpose_212")];
            tensor<fp16, [1, 110, 128]> input_13_cast_fp16 = reshape(shape = var_156, x = var_155_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<fp16, [768, 128]> layers_0_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5538240)))];
            tensor<fp16, [1, 110, 768]> linear_4_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_cross_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [1]> x_5_axes_0 = const()[name = tensor<string, []>("x_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_0_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5734912)))];
            tensor<fp16, [1, 110, 768]> x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, epsilon = var_61_to_fp16, gamma = layers_0_norm_ff_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
            tensor<int32, [3]> input_17_perm_0 = const()[name = tensor<string, []>("input_17_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_0_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5736512)))];
            tensor<fp16, [1, 768, 110]> input_17_cast_fp16 = transpose(perm = input_17_perm_0, x = x_5_cast_fp16)[name = tensor<string, []>("transpose_211")];
            tensor<fp16, [1, 3072, 110]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_0_ffn_conv1_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> input_21_mode_0 = const()[name = tensor<string, []>("input_21_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<string, []> x_7_pad_type_0 = const()[name = tensor<string, []>("x_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_7_strides_0 = const()[name = tensor<string, []>("x_7_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_7_pad_0 = const()[name = tensor<string, []>("x_7_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_7_dilations_0 = const()[name = tensor<string, []>("x_7_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_7_groups_0 = const()[name = tensor<string, []>("x_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_0_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10455168)))];
            tensor<fp16, [1, 768, 110]> x_7_cast_fp16 = conv(dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = layers_0_ffn_conv2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<int32, [3]> x_9_perm_0 = const()[name = tensor<string, []>("x_9_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_210")];
            tensor<fp16, [1, 110, 768]> input_23_cast_fp16 = add(x = input_15_cast_fp16, y = x_9_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [8]> k_padded_1_pad_0 = const()[name = tensor<string, []>("k_padded_1_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_1_mode_0 = const()[name = tensor<string, []>("k_padded_1_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_7_to_fp16 = const()[name = tensor<string, []>("const_7_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_1_cast_fp16 = pad(constant_val = const_7_to_fp16, mode = k_padded_1_mode_0, pad = k_padded_1_pad_0, x = k_1_cast_fp16)[name = tensor<string, []>("k_padded_1_cast_fp16")];
            tensor<int32, [8]> v_padded_1_pad_0 = const()[name = tensor<string, []>("v_padded_1_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_1_mode_0 = const()[name = tensor<string, []>("v_padded_1_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_8_to_fp16 = const()[name = tensor<string, []>("const_8_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_1_cast_fp16 = pad(constant_val = const_8_to_fp16, mode = v_padded_1_mode_0, pad = v_padded_1_pad_0, x = v_1_cast_fp16)[name = tensor<string, []>("v_padded_1_cast_fp16")];
            tensor<int32, []> var_208_axis_0 = const()[name = tensor<string, []>("op_208_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_208 = stack(axis = var_208_axis_0, values = (k_padded_1_cast_fp16, v_padded_1_cast_fp16))[name = tensor<string, []>("op_208_cast_fp16")];
            tensor<int32, []> var_220 = const()[name = tensor<string, []>("op_220"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_11_axes_0 = const()[name = tensor<string, []>("x_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_1_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15173824)))];
            tensor<fp16, []> var_227_to_fp16 = const()[name = tensor<string, []>("op_227_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, epsilon = var_227_to_fp16, gamma = layers_1_norm_sa_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_1_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15175424)))];
            tensor<fp16, [1, 110, 2304]> linear_5_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_qkv_proj_weight_to_fp16, x = x_11_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [5]> var_245 = const()[name = tensor<string, []>("op_245"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_3_cast_fp16 = reshape(shape = var_245, x = linear_5_cast_fp16)[name = tensor<string, []>("qkv_3_cast_fp16")];
            tensor<int32, [5]> q_7_begin_0 = const()[name = tensor<string, []>("q_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_7_end_0 = const()[name = tensor<string, []>("q_7_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_7_end_mask_0 = const()[name = tensor<string, []>("q_7_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_7_squeeze_mask_0 = const()[name = tensor<string, []>("q_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
            tensor<int32, [5]> k_5_begin_0 = const()[name = tensor<string, []>("k_5_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_5_end_0 = const()[name = tensor<string, []>("k_5_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_5_end_mask_0 = const()[name = tensor<string, []>("k_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_5_squeeze_mask_0 = const()[name = tensor<string, []>("k_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
            tensor<int32, [5]> v_5_begin_0 = const()[name = tensor<string, []>("v_5_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_5_end_0 = const()[name = tensor<string, []>("v_5_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_5_end_mask_0 = const()[name = tensor<string, []>("v_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_5_squeeze_mask_0 = const()[name = tensor<string, []>("v_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
            tensor<int32, [4]> v_t_3_perm_0 = const()[name = tensor<string, []>("v_t_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_260_transpose_x_0 = const()[name = tensor<string, []>("op_260_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_260_transpose_y_0 = const()[name = tensor<string, []>("op_260_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = tensor<string, []>("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_71 = transpose(perm = transpose_71_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_207")];
            tensor<fp16, [1, 12, 110, 64]> transpose_70 = transpose(perm = transpose_70_perm_0, x = q_7_cast_fp16)[name = tensor<string, []>("transpose_208")];
            tensor<fp16, [1, 12, 110, 110]> var_260_cast_fp16 = matmul(transpose_x = var_260_transpose_x_0, transpose_y = var_260_transpose_y_0, x = transpose_70, y = transpose_71)[name = tensor<string, []>("op_260_cast_fp16")];
            tensor<fp16, []> var_261_to_fp16 = const()[name = tensor<string, []>("op_261_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_9_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<string, []> attn_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_9_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_9_cast_fp16_to_fp32 = cast(dtype = attn_9_cast_fp16_to_fp32_dtype_0, x = attn_9_cast_fp16)[name = tensor<string, []>("cast_91")];
            tensor<fp32, [1, 12, 110, 110]> input_25 = add(x = attn_9_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_25")];
            tensor<string, []> input_25_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_25_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_25_to_fp16 = cast(dtype = input_25_to_fp16_dtype_0, x = input_25)[name = tensor<string, []>("cast_90")];
            tensor<fp16, [1, 12, 110, 110]> attn_11_cast_fp16 = softmax(axis = var_220, x = input_25_to_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
            tensor<bool, []> out_5_transpose_x_0 = const()[name = tensor<string, []>("out_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_5_transpose_y_0 = const()[name = tensor<string, []>("out_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_3_cast_fp16 = transpose(perm = v_t_3_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_209")];
            tensor<fp16, [1, 12, 110, 64]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_11_cast_fp16, y = v_t_3_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<int32, [4]> var_272_perm_0 = const()[name = tensor<string, []>("op_272_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_273 = const()[name = tensor<string, []>("op_273"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_272_cast_fp16 = transpose(perm = var_272_perm_0, x = out_5_cast_fp16)[name = tensor<string, []>("transpose_206")];
            tensor<fp16, [1, 110, 768]> input_27_cast_fp16 = reshape(shape = var_273, x = var_272_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<fp16, [768, 768]> 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, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18714432)))];
            tensor<fp16, [1, 110, 768]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_29_cast_fp16 = add(x = input_23_cast_fp16, y = linear_6_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<int32, [1]> x_13_axes_0 = const()[name = tensor<string, []>("x_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_1_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19894144)))];
            tensor<fp16, [1, 110, 768]> x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, epsilon = var_227_to_fp16, gamma = layers_1_norm_xa_query_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<int32, [1]> memory_3_axes_0 = const()[name = tensor<string, []>("memory_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_1_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19895744)))];
            tensor<fp16, [1, 256, 768]> memory_3_cast_fp16 = layer_norm(axes = memory_3_axes_0, epsilon = var_227_to_fp16, gamma = layers_1_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_3_cast_fp16")];
            tensor<fp16, [128, 768]> layers_1_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19897344)))];
            tensor<fp16, [1, 110, 128]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_cross_attn_q_proj_weight_to_fp16, x = x_13_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<int32, [4]> var_296 = const()[name = tensor<string, []>("op_296"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_297_cast_fp16 = reshape(shape = var_296, x = linear_7_cast_fp16)[name = tensor<string, []>("op_297_cast_fp16")];
            tensor<fp16, [256, 768]> layers_1_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20094016)))];
            tensor<fp16, [1, 256, 256]> linear_8_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_1_cross_attn_kv_proj_weight_to_fp16, x = memory_3_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<int32, [5]> var_301 = const()[name = tensor<string, []>("op_301"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_3_cast_fp16 = reshape(shape = var_301, x = linear_8_cast_fp16)[name = tensor<string, []>("kv_3_cast_fp16")];
            tensor<int32, [5]> var_305_begin_0 = const()[name = tensor<string, []>("op_305_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_305_end_0 = const()[name = tensor<string, []>("op_305_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_305_end_mask_0 = const()[name = tensor<string, []>("op_305_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_305_squeeze_mask_0 = const()[name = tensor<string, []>("op_305_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_305_cast_fp16 = slice_by_index(begin = var_305_begin_0, end = var_305_end_0, end_mask = var_305_end_mask_0, squeeze_mask = var_305_squeeze_mask_0, x = kv_3_cast_fp16)[name = tensor<string, []>("op_305_cast_fp16")];
            tensor<int32, [5]> var_309_begin_0 = const()[name = tensor<string, []>("op_309_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_309_end_0 = const()[name = tensor<string, []>("op_309_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_309_end_mask_0 = const()[name = tensor<string, []>("op_309_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_309_squeeze_mask_0 = const()[name = tensor<string, []>("op_309_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_309_cast_fp16 = slice_by_index(begin = var_309_begin_0, end = var_309_end_0, end_mask = var_309_end_mask_0, squeeze_mask = var_309_squeeze_mask_0, x = kv_3_cast_fp16)[name = tensor<string, []>("op_309_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = tensor<string, []>("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_312_transpose_x_0 = const()[name = tensor<string, []>("op_312_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_312_transpose_y_0 = const()[name = tensor<string, []>("op_312_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = tensor<string, []>("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = tensor<string, []>("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_305_cast_fp16)[name = tensor<string, []>("transpose_203")];
            tensor<fp16, [1, 1, 110, 128]> transpose_72 = transpose(perm = transpose_72_perm_0, x = var_297_cast_fp16)[name = tensor<string, []>("transpose_204")];
            tensor<fp16, [1, 1, 110, 256]> var_312_cast_fp16 = matmul(transpose_x = var_312_transpose_x_0, transpose_y = var_312_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor<string, []>("op_312_cast_fp16")];
            tensor<fp16, []> var_313_to_fp16 = const()[name = tensor<string, []>("op_313_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_13_cast_fp16 = mul(x = var_312_cast_fp16, y = var_313_to_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<fp16, []> var_211_to_fp16 = const()[name = tensor<string, []>("op_211_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_31_cast_fp16 = select(a = var_211_to_fp16, b = attn_13_cast_fp16, cond = var_151)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_15_cast_fp16 = softmax(axis = var_220, x = input_31_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
            tensor<bool, []> out_7_transpose_x_0 = const()[name = tensor<string, []>("out_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_7_transpose_y_0 = const()[name = tensor<string, []>("out_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_309_cast_fp16)[name = tensor<string, []>("transpose_205")];
            tensor<fp16, [1, 1, 110, 128]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_15_cast_fp16, y = v_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<int32, [4]> var_321_perm_0 = const()[name = tensor<string, []>("op_321_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_322 = const()[name = tensor<string, []>("op_322"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_321_cast_fp16 = transpose(perm = var_321_perm_0, x = out_7_cast_fp16)[name = tensor<string, []>("transpose_202")];
            tensor<fp16, [1, 110, 128]> input_33_cast_fp16 = reshape(shape = var_322, x = var_321_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<fp16, [768, 128]> layers_1_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20487296)))];
            tensor<fp16, [1, 110, 768]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_cross_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_35_cast_fp16 = add(x = input_29_cast_fp16, y = linear_9_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<int32, [1]> x_15_axes_0 = const()[name = tensor<string, []>("x_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_1_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20683968)))];
            tensor<fp16, [1, 110, 768]> x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, epsilon = var_227_to_fp16, gamma = layers_1_norm_ff_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
            tensor<int32, [3]> input_37_perm_0 = const()[name = tensor<string, []>("input_37_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_39_strides_0 = const()[name = tensor<string, []>("input_39_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_39_dilations_0 = const()[name = tensor<string, []>("input_39_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_39_groups_0 = const()[name = tensor<string, []>("input_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_1_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20685568)))];
            tensor<fp16, [1, 768, 110]> input_37_cast_fp16 = transpose(perm = input_37_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_201")];
            tensor<fp16, [1, 3072, 110]> input_39_cast_fp16 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_1_ffn_conv1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<string, []> input_41_mode_0 = const()[name = tensor<string, []>("input_41_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<string, []> x_17_pad_type_0 = const()[name = tensor<string, []>("x_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_17_strides_0 = const()[name = tensor<string, []>("x_17_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_17_pad_0 = const()[name = tensor<string, []>("x_17_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_17_dilations_0 = const()[name = tensor<string, []>("x_17_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_17_groups_0 = const()[name = tensor<string, []>("x_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_1_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25404224)))];
            tensor<fp16, [1, 768, 110]> x_17_cast_fp16 = conv(dilations = x_17_dilations_0, groups = x_17_groups_0, pad = x_17_pad_0, pad_type = x_17_pad_type_0, strides = x_17_strides_0, weight = layers_1_ffn_conv2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<int32, [3]> x_19_perm_0 = const()[name = tensor<string, []>("x_19_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_19_cast_fp16 = transpose(perm = x_19_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_200")];
            tensor<fp16, [1, 110, 768]> input_43_cast_fp16 = add(x = input_35_cast_fp16, y = x_19_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<int32, [8]> k_padded_3_pad_0 = const()[name = tensor<string, []>("k_padded_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_3_mode_0 = const()[name = tensor<string, []>("k_padded_3_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_14_to_fp16 = const()[name = tensor<string, []>("const_14_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_3_cast_fp16 = pad(constant_val = const_14_to_fp16, mode = k_padded_3_mode_0, pad = k_padded_3_pad_0, x = k_5_cast_fp16)[name = tensor<string, []>("k_padded_3_cast_fp16")];
            tensor<int32, [8]> v_padded_3_pad_0 = const()[name = tensor<string, []>("v_padded_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_3_mode_0 = const()[name = tensor<string, []>("v_padded_3_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_15_to_fp16 = const()[name = tensor<string, []>("const_15_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_3_cast_fp16 = pad(constant_val = const_15_to_fp16, mode = v_padded_3_mode_0, pad = v_padded_3_pad_0, x = v_5_cast_fp16)[name = tensor<string, []>("v_padded_3_cast_fp16")];
            tensor<int32, []> var_374_axis_0 = const()[name = tensor<string, []>("op_374_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_374 = stack(axis = var_374_axis_0, values = (k_padded_3_cast_fp16, v_padded_3_cast_fp16))[name = tensor<string, []>("op_374_cast_fp16")];
            tensor<int32, []> var_386 = const()[name = tensor<string, []>("op_386"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_21_axes_0 = const()[name = tensor<string, []>("x_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_2_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30122880)))];
            tensor<fp16, []> var_393_to_fp16 = const()[name = tensor<string, []>("op_393_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, epsilon = var_393_to_fp16, gamma = layers_2_norm_sa_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_2_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30124480)))];
            tensor<fp16, [1, 110, 2304]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_qkv_proj_weight_to_fp16, x = x_21_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<int32, [5]> var_411 = const()[name = tensor<string, []>("op_411"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_5_cast_fp16 = reshape(shape = var_411, x = linear_10_cast_fp16)[name = tensor<string, []>("qkv_5_cast_fp16")];
            tensor<int32, [5]> q_13_begin_0 = const()[name = tensor<string, []>("q_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_13_end_0 = const()[name = tensor<string, []>("q_13_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_13_end_mask_0 = const()[name = tensor<string, []>("q_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_13_squeeze_mask_0 = const()[name = tensor<string, []>("q_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
            tensor<int32, [5]> k_9_begin_0 = const()[name = tensor<string, []>("k_9_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_9_end_0 = const()[name = tensor<string, []>("k_9_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_9_end_mask_0 = const()[name = tensor<string, []>("k_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_9_squeeze_mask_0 = const()[name = tensor<string, []>("k_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
            tensor<int32, [5]> v_9_begin_0 = const()[name = tensor<string, []>("v_9_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_9_end_0 = const()[name = tensor<string, []>("v_9_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_9_end_mask_0 = const()[name = tensor<string, []>("v_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_9_squeeze_mask_0 = const()[name = tensor<string, []>("v_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
            tensor<int32, [4]> v_t_5_perm_0 = const()[name = tensor<string, []>("v_t_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_426_transpose_x_0 = const()[name = tensor<string, []>("op_426_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_426_transpose_y_0 = const()[name = tensor<string, []>("op_426_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = tensor<string, []>("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = tensor<string, []>("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_75 = transpose(perm = transpose_75_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_197")];
            tensor<fp16, [1, 12, 110, 64]> transpose_74 = transpose(perm = transpose_74_perm_0, x = q_13_cast_fp16)[name = tensor<string, []>("transpose_198")];
            tensor<fp16, [1, 12, 110, 110]> var_426_cast_fp16 = matmul(transpose_x = var_426_transpose_x_0, transpose_y = var_426_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor<string, []>("op_426_cast_fp16")];
            tensor<fp16, []> var_427_to_fp16 = const()[name = tensor<string, []>("op_427_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_17_cast_fp16 = mul(x = var_426_cast_fp16, y = var_427_to_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<string, []> attn_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_17_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_17_cast_fp16_to_fp32 = cast(dtype = attn_17_cast_fp16_to_fp32_dtype_0, x = attn_17_cast_fp16)[name = tensor<string, []>("cast_89")];
            tensor<fp32, [1, 12, 110, 110]> input_45 = add(x = attn_17_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_45")];
            tensor<string, []> input_45_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_45_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_45_to_fp16 = cast(dtype = input_45_to_fp16_dtype_0, x = input_45)[name = tensor<string, []>("cast_88")];
            tensor<fp16, [1, 12, 110, 110]> attn_19_cast_fp16 = softmax(axis = var_386, x = input_45_to_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
            tensor<bool, []> out_9_transpose_x_0 = const()[name = tensor<string, []>("out_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_9_transpose_y_0 = const()[name = tensor<string, []>("out_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_5_cast_fp16 = transpose(perm = v_t_5_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_199")];
            tensor<fp16, [1, 12, 110, 64]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_19_cast_fp16, y = v_t_5_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<int32, [4]> var_438_perm_0 = const()[name = tensor<string, []>("op_438_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_439 = const()[name = tensor<string, []>("op_439"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_438_cast_fp16 = transpose(perm = var_438_perm_0, x = out_9_cast_fp16)[name = tensor<string, []>("transpose_196")];
            tensor<fp16, [1, 110, 768]> input_47_cast_fp16 = reshape(shape = var_439, x = var_438_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<fp16, [768, 768]> 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, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33663488)))];
            tensor<fp16, [1, 110, 768]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_2_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34843200)))];
            tensor<fp16, [1, 110, 768]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, epsilon = var_393_to_fp16, gamma = layers_2_norm_xa_query_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<int32, [1]> memory_5_axes_0 = const()[name = tensor<string, []>("memory_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_2_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34844800)))];
            tensor<fp16, [1, 256, 768]> memory_5_cast_fp16 = layer_norm(axes = memory_5_axes_0, epsilon = var_393_to_fp16, gamma = layers_2_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_5_cast_fp16")];
            tensor<fp16, [128, 768]> layers_2_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34846400)))];
            tensor<fp16, [1, 110, 128]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_cross_attn_q_proj_weight_to_fp16, x = x_23_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<int32, [4]> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_463_cast_fp16 = reshape(shape = var_462, x = linear_12_cast_fp16)[name = tensor<string, []>("op_463_cast_fp16")];
            tensor<fp16, [256, 768]> layers_2_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35043072)))];
            tensor<fp16, [1, 256, 256]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_2_cross_attn_kv_proj_weight_to_fp16, x = memory_5_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<int32, [5]> var_467 = const()[name = tensor<string, []>("op_467"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_5_cast_fp16 = reshape(shape = var_467, x = linear_13_cast_fp16)[name = tensor<string, []>("kv_5_cast_fp16")];
            tensor<int32, [5]> var_471_begin_0 = const()[name = tensor<string, []>("op_471_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_471_end_0 = const()[name = tensor<string, []>("op_471_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_471_end_mask_0 = const()[name = tensor<string, []>("op_471_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_471_squeeze_mask_0 = const()[name = tensor<string, []>("op_471_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_471_cast_fp16 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, squeeze_mask = var_471_squeeze_mask_0, x = kv_5_cast_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
            tensor<int32, [5]> var_475_begin_0 = const()[name = tensor<string, []>("op_475_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_475_end_0 = const()[name = tensor<string, []>("op_475_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_475_end_mask_0 = const()[name = tensor<string, []>("op_475_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_475_squeeze_mask_0 = const()[name = tensor<string, []>("op_475_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_475_cast_fp16 = slice_by_index(begin = var_475_begin_0, end = var_475_end_0, end_mask = var_475_end_mask_0, squeeze_mask = var_475_squeeze_mask_0, x = kv_5_cast_fp16)[name = tensor<string, []>("op_475_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = tensor<string, []>("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_478_transpose_x_0 = const()[name = tensor<string, []>("op_478_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_478_transpose_y_0 = const()[name = tensor<string, []>("op_478_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = tensor<string, []>("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_77_perm_0 = const()[name = tensor<string, []>("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_471_cast_fp16)[name = tensor<string, []>("transpose_193")];
            tensor<fp16, [1, 1, 110, 128]> transpose_76 = transpose(perm = transpose_76_perm_0, x = var_463_cast_fp16)[name = tensor<string, []>("transpose_194")];
            tensor<fp16, [1, 1, 110, 256]> var_478_cast_fp16 = matmul(transpose_x = var_478_transpose_x_0, transpose_y = var_478_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor<string, []>("op_478_cast_fp16")];
            tensor<fp16, []> var_479_to_fp16 = const()[name = tensor<string, []>("op_479_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_21_cast_fp16 = mul(x = var_478_cast_fp16, y = var_479_to_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<fp16, []> var_377_to_fp16 = const()[name = tensor<string, []>("op_377_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_51_cast_fp16 = select(a = var_377_to_fp16, b = attn_21_cast_fp16, cond = var_151)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_23_cast_fp16 = softmax(axis = var_386, x = input_51_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
            tensor<bool, []> out_11_transpose_x_0 = const()[name = tensor<string, []>("out_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_11_transpose_y_0 = const()[name = tensor<string, []>("out_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_475_cast_fp16)[name = tensor<string, []>("transpose_195")];
            tensor<fp16, [1, 1, 110, 128]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_23_cast_fp16, y = v_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<int32, [4]> var_487_perm_0 = const()[name = tensor<string, []>("op_487_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_488 = const()[name = tensor<string, []>("op_488"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_487_cast_fp16 = transpose(perm = var_487_perm_0, x = out_11_cast_fp16)[name = tensor<string, []>("transpose_192")];
            tensor<fp16, [1, 110, 128]> input_53_cast_fp16 = reshape(shape = var_488, x = var_487_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<fp16, [768, 128]> layers_2_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35436352)))];
            tensor<fp16, [1, 110, 768]> linear_14_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_cross_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_14_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_2_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35633024)))];
            tensor<fp16, [1, 110, 768]> x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, epsilon = var_393_to_fp16, gamma = layers_2_norm_ff_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_59_pad_type_0 = const()[name = tensor<string, []>("input_59_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_59_strides_0 = const()[name = tensor<string, []>("input_59_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_59_pad_0 = const()[name = tensor<string, []>("input_59_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_59_dilations_0 = const()[name = tensor<string, []>("input_59_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_59_groups_0 = const()[name = tensor<string, []>("input_59_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_2_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35634624)))];
            tensor<fp16, [1, 768, 110]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_191")];
            tensor<fp16, [1, 3072, 110]> input_59_cast_fp16 = conv(dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = layers_2_ffn_conv1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<string, []> input_61_mode_0 = const()[name = tensor<string, []>("input_61_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<string, []> x_27_pad_type_0 = const()[name = tensor<string, []>("x_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_27_strides_0 = const()[name = tensor<string, []>("x_27_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_27_pad_0 = const()[name = tensor<string, []>("x_27_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_27_dilations_0 = const()[name = tensor<string, []>("x_27_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_27_groups_0 = const()[name = tensor<string, []>("x_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_2_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40353280)))];
            tensor<fp16, [1, 768, 110]> x_27_cast_fp16 = conv(dilations = x_27_dilations_0, groups = x_27_groups_0, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = x_27_strides_0, weight = layers_2_ffn_conv2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<int32, [3]> x_29_perm_0 = const()[name = tensor<string, []>("x_29_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = x_27_cast_fp16)[name = tensor<string, []>("transpose_190")];
            tensor<fp16, [1, 110, 768]> input_63_cast_fp16 = add(x = input_55_cast_fp16, y = x_29_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<int32, [8]> k_padded_5_pad_0 = const()[name = tensor<string, []>("k_padded_5_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_5_mode_0 = const()[name = tensor<string, []>("k_padded_5_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_21_to_fp16 = const()[name = tensor<string, []>("const_21_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_5_cast_fp16 = pad(constant_val = const_21_to_fp16, mode = k_padded_5_mode_0, pad = k_padded_5_pad_0, x = k_9_cast_fp16)[name = tensor<string, []>("k_padded_5_cast_fp16")];
            tensor<int32, [8]> v_padded_5_pad_0 = const()[name = tensor<string, []>("v_padded_5_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_5_mode_0 = const()[name = tensor<string, []>("v_padded_5_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_22_to_fp16 = const()[name = tensor<string, []>("const_22_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_5_cast_fp16 = pad(constant_val = const_22_to_fp16, mode = v_padded_5_mode_0, pad = v_padded_5_pad_0, x = v_9_cast_fp16)[name = tensor<string, []>("v_padded_5_cast_fp16")];
            tensor<int32, []> var_540_axis_0 = const()[name = tensor<string, []>("op_540_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_540 = stack(axis = var_540_axis_0, values = (k_padded_5_cast_fp16, v_padded_5_cast_fp16))[name = tensor<string, []>("op_540_cast_fp16")];
            tensor<int32, []> var_552 = const()[name = tensor<string, []>("op_552"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_31_axes_0 = const()[name = tensor<string, []>("x_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_3_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45071936)))];
            tensor<fp16, []> var_559_to_fp16 = const()[name = tensor<string, []>("op_559_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, epsilon = var_559_to_fp16, gamma = layers_3_norm_sa_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_3_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45073536)))];
            tensor<fp16, [1, 110, 2304]> linear_15_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_qkv_proj_weight_to_fp16, x = x_31_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<int32, [5]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_7_cast_fp16 = reshape(shape = var_577, x = linear_15_cast_fp16)[name = tensor<string, []>("qkv_7_cast_fp16")];
            tensor<int32, [5]> q_19_begin_0 = const()[name = tensor<string, []>("q_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_19_end_0 = const()[name = tensor<string, []>("q_19_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_19_end_mask_0 = const()[name = tensor<string, []>("q_19_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_19_squeeze_mask_0 = const()[name = tensor<string, []>("q_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
            tensor<int32, [5]> k_13_begin_0 = const()[name = tensor<string, []>("k_13_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_13_end_0 = const()[name = tensor<string, []>("k_13_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_13_end_mask_0 = const()[name = tensor<string, []>("k_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
            tensor<int32, [5]> v_13_begin_0 = const()[name = tensor<string, []>("v_13_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_13_end_0 = const()[name = tensor<string, []>("v_13_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_13_end_mask_0 = const()[name = tensor<string, []>("v_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_13_squeeze_mask_0 = const()[name = tensor<string, []>("v_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
            tensor<int32, [4]> v_t_7_perm_0 = const()[name = tensor<string, []>("v_t_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_592_transpose_x_0 = const()[name = tensor<string, []>("op_592_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_592_transpose_y_0 = const()[name = tensor<string, []>("op_592_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = tensor<string, []>("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = tensor<string, []>("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_79 = transpose(perm = transpose_79_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_187")];
            tensor<fp16, [1, 12, 110, 64]> transpose_78 = transpose(perm = transpose_78_perm_0, x = q_19_cast_fp16)[name = tensor<string, []>("transpose_188")];
            tensor<fp16, [1, 12, 110, 110]> var_592_cast_fp16 = matmul(transpose_x = var_592_transpose_x_0, transpose_y = var_592_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor<string, []>("op_592_cast_fp16")];
            tensor<fp16, []> var_593_to_fp16 = const()[name = tensor<string, []>("op_593_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_25_cast_fp16 = mul(x = var_592_cast_fp16, y = var_593_to_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
            tensor<string, []> attn_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_25_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_25_cast_fp16_to_fp32 = cast(dtype = attn_25_cast_fp16_to_fp32_dtype_0, x = attn_25_cast_fp16)[name = tensor<string, []>("cast_87")];
            tensor<fp32, [1, 12, 110, 110]> input_65 = add(x = attn_25_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_65")];
            tensor<string, []> input_65_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_65_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_65_to_fp16 = cast(dtype = input_65_to_fp16_dtype_0, x = input_65)[name = tensor<string, []>("cast_86")];
            tensor<fp16, [1, 12, 110, 110]> attn_27_cast_fp16 = softmax(axis = var_552, x = input_65_to_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
            tensor<bool, []> out_13_transpose_x_0 = const()[name = tensor<string, []>("out_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_13_transpose_y_0 = const()[name = tensor<string, []>("out_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_7_cast_fp16 = transpose(perm = v_t_7_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_189")];
            tensor<fp16, [1, 12, 110, 64]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_27_cast_fp16, y = v_t_7_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<int32, [4]> var_604_perm_0 = const()[name = tensor<string, []>("op_604_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_605 = const()[name = tensor<string, []>("op_605"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_604_cast_fp16 = transpose(perm = var_604_perm_0, x = out_13_cast_fp16)[name = tensor<string, []>("transpose_186")];
            tensor<fp16, [1, 110, 768]> input_67_cast_fp16 = reshape(shape = var_605, x = var_604_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<fp16, [768, 768]> 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, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48612544)))];
            tensor<fp16, [1, 110, 768]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<int32, [1]> x_33_axes_0 = const()[name = tensor<string, []>("x_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_3_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49792256)))];
            tensor<fp16, [1, 110, 768]> x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, epsilon = var_559_to_fp16, gamma = layers_3_norm_xa_query_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
            tensor<int32, [1]> memory_7_axes_0 = const()[name = tensor<string, []>("memory_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_3_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49793856)))];
            tensor<fp16, [1, 256, 768]> memory_7_cast_fp16 = layer_norm(axes = memory_7_axes_0, epsilon = var_559_to_fp16, gamma = layers_3_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_7_cast_fp16")];
            tensor<fp16, [128, 768]> layers_3_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49795456)))];
            tensor<fp16, [1, 110, 128]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_cross_attn_q_proj_weight_to_fp16, x = x_33_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<int32, [4]> var_628 = const()[name = tensor<string, []>("op_628"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_629_cast_fp16 = reshape(shape = var_628, x = linear_17_cast_fp16)[name = tensor<string, []>("op_629_cast_fp16")];
            tensor<fp16, [256, 768]> layers_3_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49992128)))];
            tensor<fp16, [1, 256, 256]> linear_18_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_3_cross_attn_kv_proj_weight_to_fp16, x = memory_7_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<int32, [5]> var_633 = const()[name = tensor<string, []>("op_633"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_7_cast_fp16 = reshape(shape = var_633, x = linear_18_cast_fp16)[name = tensor<string, []>("kv_7_cast_fp16")];
            tensor<int32, [5]> var_637_begin_0 = const()[name = tensor<string, []>("op_637_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_637_end_0 = const()[name = tensor<string, []>("op_637_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_637_end_mask_0 = const()[name = tensor<string, []>("op_637_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_637_squeeze_mask_0 = const()[name = tensor<string, []>("op_637_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_637_cast_fp16 = slice_by_index(begin = var_637_begin_0, end = var_637_end_0, end_mask = var_637_end_mask_0, squeeze_mask = var_637_squeeze_mask_0, x = kv_7_cast_fp16)[name = tensor<string, []>("op_637_cast_fp16")];
            tensor<int32, [5]> var_641_begin_0 = const()[name = tensor<string, []>("op_641_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_641_end_0 = const()[name = tensor<string, []>("op_641_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_641_end_mask_0 = const()[name = tensor<string, []>("op_641_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_641_squeeze_mask_0 = const()[name = tensor<string, []>("op_641_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_641_cast_fp16 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, squeeze_mask = var_641_squeeze_mask_0, x = kv_7_cast_fp16)[name = tensor<string, []>("op_641_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = tensor<string, []>("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_644_transpose_x_0 = const()[name = tensor<string, []>("op_644_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_644_transpose_y_0 = const()[name = tensor<string, []>("op_644_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = tensor<string, []>("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = tensor<string, []>("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_637_cast_fp16)[name = tensor<string, []>("transpose_183")];
            tensor<fp16, [1, 1, 110, 128]> transpose_80 = transpose(perm = transpose_80_perm_0, x = var_629_cast_fp16)[name = tensor<string, []>("transpose_184")];
            tensor<fp16, [1, 1, 110, 256]> var_644_cast_fp16 = matmul(transpose_x = var_644_transpose_x_0, transpose_y = var_644_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor<string, []>("op_644_cast_fp16")];
            tensor<fp16, []> var_645_to_fp16 = const()[name = tensor<string, []>("op_645_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_29_cast_fp16 = mul(x = var_644_cast_fp16, y = var_645_to_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
            tensor<fp16, []> var_543_to_fp16 = const()[name = tensor<string, []>("op_543_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_71_cast_fp16 = select(a = var_543_to_fp16, b = attn_29_cast_fp16, cond = var_151)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_31_cast_fp16 = softmax(axis = var_552, x = input_71_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
            tensor<bool, []> out_15_transpose_x_0 = const()[name = tensor<string, []>("out_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_15_transpose_y_0 = const()[name = tensor<string, []>("out_15_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_641_cast_fp16)[name = tensor<string, []>("transpose_185")];
            tensor<fp16, [1, 1, 110, 128]> out_15_cast_fp16 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_31_cast_fp16, y = v_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<int32, [4]> var_653_perm_0 = const()[name = tensor<string, []>("op_653_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_654 = const()[name = tensor<string, []>("op_654"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_653_cast_fp16 = transpose(perm = var_653_perm_0, x = out_15_cast_fp16)[name = tensor<string, []>("transpose_182")];
            tensor<fp16, [1, 110, 128]> input_73_cast_fp16 = reshape(shape = var_654, x = var_653_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<fp16, [768, 128]> layers_3_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50385408)))];
            tensor<fp16, [1, 110, 768]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_cross_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<int32, [1]> x_35_axes_0 = const()[name = tensor<string, []>("x_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_3_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50582080)))];
            tensor<fp16, [1, 110, 768]> x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, epsilon = var_559_to_fp16, gamma = layers_3_norm_ff_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<int32, [3]> input_77_perm_0 = const()[name = tensor<string, []>("input_77_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_79_pad_type_0 = const()[name = tensor<string, []>("input_79_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_79_strides_0 = const()[name = tensor<string, []>("input_79_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_79_pad_0 = const()[name = tensor<string, []>("input_79_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_79_dilations_0 = const()[name = tensor<string, []>("input_79_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_79_groups_0 = const()[name = tensor<string, []>("input_79_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_3_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50583680)))];
            tensor<fp16, [1, 768, 110]> input_77_cast_fp16 = transpose(perm = input_77_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_181")];
            tensor<fp16, [1, 3072, 110]> input_79_cast_fp16 = conv(dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = layers_3_ffn_conv1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<string, []> input_81_mode_0 = const()[name = tensor<string, []>("input_81_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<string, []> x_37_pad_type_0 = const()[name = tensor<string, []>("x_37_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_37_strides_0 = const()[name = tensor<string, []>("x_37_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_37_pad_0 = const()[name = tensor<string, []>("x_37_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_37_dilations_0 = const()[name = tensor<string, []>("x_37_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_37_groups_0 = const()[name = tensor<string, []>("x_37_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_3_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55302336)))];
            tensor<fp16, [1, 768, 110]> x_37_cast_fp16 = conv(dilations = x_37_dilations_0, groups = x_37_groups_0, pad = x_37_pad_0, pad_type = x_37_pad_type_0, strides = x_37_strides_0, weight = layers_3_ffn_conv2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<int32, [3]> x_39_perm_0 = const()[name = tensor<string, []>("x_39_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = x_37_cast_fp16)[name = tensor<string, []>("transpose_180")];
            tensor<fp16, [1, 110, 768]> input_83_cast_fp16 = add(x = input_75_cast_fp16, y = x_39_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<int32, [8]> k_padded_7_pad_0 = const()[name = tensor<string, []>("k_padded_7_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_7_mode_0 = const()[name = tensor<string, []>("k_padded_7_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_28_to_fp16 = const()[name = tensor<string, []>("const_28_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_7_cast_fp16 = pad(constant_val = const_28_to_fp16, mode = k_padded_7_mode_0, pad = k_padded_7_pad_0, x = k_13_cast_fp16)[name = tensor<string, []>("k_padded_7_cast_fp16")];
            tensor<int32, [8]> v_padded_7_pad_0 = const()[name = tensor<string, []>("v_padded_7_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_7_mode_0 = const()[name = tensor<string, []>("v_padded_7_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_29_to_fp16 = const()[name = tensor<string, []>("const_29_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_7_cast_fp16 = pad(constant_val = const_29_to_fp16, mode = v_padded_7_mode_0, pad = v_padded_7_pad_0, x = v_13_cast_fp16)[name = tensor<string, []>("v_padded_7_cast_fp16")];
            tensor<int32, []> var_706_axis_0 = const()[name = tensor<string, []>("op_706_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_706 = stack(axis = var_706_axis_0, values = (k_padded_7_cast_fp16, v_padded_7_cast_fp16))[name = tensor<string, []>("op_706_cast_fp16")];
            tensor<int32, []> var_718 = const()[name = tensor<string, []>("op_718"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_41_axes_0 = const()[name = tensor<string, []>("x_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_4_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60020992)))];
            tensor<fp16, []> var_725_to_fp16 = const()[name = tensor<string, []>("op_725_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, epsilon = var_725_to_fp16, gamma = layers_4_norm_sa_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_4_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60022592)))];
            tensor<fp16, [1, 110, 2304]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_qkv_proj_weight_to_fp16, x = x_41_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<int32, [5]> var_743 = const()[name = tensor<string, []>("op_743"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_9_cast_fp16 = reshape(shape = var_743, x = linear_20_cast_fp16)[name = tensor<string, []>("qkv_9_cast_fp16")];
            tensor<int32, [5]> q_25_begin_0 = const()[name = tensor<string, []>("q_25_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_25_end_0 = const()[name = tensor<string, []>("q_25_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_25_end_mask_0 = const()[name = tensor<string, []>("q_25_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_25_squeeze_mask_0 = const()[name = tensor<string, []>("q_25_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
            tensor<int32, [5]> k_17_begin_0 = const()[name = tensor<string, []>("k_17_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_17_end_0 = const()[name = tensor<string, []>("k_17_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_17_end_mask_0 = const()[name = tensor<string, []>("k_17_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_17_squeeze_mask_0 = const()[name = tensor<string, []>("k_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
            tensor<int32, [5]> v_17_begin_0 = const()[name = tensor<string, []>("v_17_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_17_end_0 = const()[name = tensor<string, []>("v_17_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_17_end_mask_0 = const()[name = tensor<string, []>("v_17_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_17_squeeze_mask_0 = const()[name = tensor<string, []>("v_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
            tensor<int32, [4]> v_t_9_perm_0 = const()[name = tensor<string, []>("v_t_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_758_transpose_x_0 = const()[name = tensor<string, []>("op_758_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_758_transpose_y_0 = const()[name = tensor<string, []>("op_758_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = tensor<string, []>("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = tensor<string, []>("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_83 = transpose(perm = transpose_83_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_177")];
            tensor<fp16, [1, 12, 110, 64]> transpose_82 = transpose(perm = transpose_82_perm_0, x = q_25_cast_fp16)[name = tensor<string, []>("transpose_178")];
            tensor<fp16, [1, 12, 110, 110]> var_758_cast_fp16 = matmul(transpose_x = var_758_transpose_x_0, transpose_y = var_758_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor<string, []>("op_758_cast_fp16")];
            tensor<fp16, []> var_759_to_fp16 = const()[name = tensor<string, []>("op_759_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_33_cast_fp16 = mul(x = var_758_cast_fp16, y = var_759_to_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
            tensor<string, []> attn_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_33_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_33_cast_fp16_to_fp32 = cast(dtype = attn_33_cast_fp16_to_fp32_dtype_0, x = attn_33_cast_fp16)[name = tensor<string, []>("cast_85")];
            tensor<fp32, [1, 12, 110, 110]> input_85 = add(x = attn_33_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_85")];
            tensor<string, []> input_85_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_85_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_85_to_fp16 = cast(dtype = input_85_to_fp16_dtype_0, x = input_85)[name = tensor<string, []>("cast_84")];
            tensor<fp16, [1, 12, 110, 110]> attn_35_cast_fp16 = softmax(axis = var_718, x = input_85_to_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
            tensor<bool, []> out_17_transpose_x_0 = const()[name = tensor<string, []>("out_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_17_transpose_y_0 = const()[name = tensor<string, []>("out_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_9_cast_fp16 = transpose(perm = v_t_9_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_179")];
            tensor<fp16, [1, 12, 110, 64]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_35_cast_fp16, y = v_t_9_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<int32, [4]> var_770_perm_0 = const()[name = tensor<string, []>("op_770_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_770_cast_fp16 = transpose(perm = var_770_perm_0, x = out_17_cast_fp16)[name = tensor<string, []>("transpose_176")];
            tensor<fp16, [1, 110, 768]> input_87_cast_fp16 = reshape(shape = var_771, x = var_770_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<fp16, [768, 768]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63561600)))];
            tensor<fp16, [1, 110, 768]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_89_cast_fp16 = add(x = input_83_cast_fp16, y = linear_21_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_4_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64741312)))];
            tensor<fp16, [1, 110, 768]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, epsilon = var_725_to_fp16, gamma = layers_4_norm_xa_query_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<int32, [1]> memory_9_axes_0 = const()[name = tensor<string, []>("memory_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_4_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64742912)))];
            tensor<fp16, [1, 256, 768]> memory_9_cast_fp16 = layer_norm(axes = memory_9_axes_0, epsilon = var_725_to_fp16, gamma = layers_4_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_9_cast_fp16")];
            tensor<fp16, [128, 768]> layers_4_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64744512)))];
            tensor<fp16, [1, 110, 128]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_cross_attn_q_proj_weight_to_fp16, x = x_43_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [4]> var_794 = const()[name = tensor<string, []>("op_794"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_795_cast_fp16 = reshape(shape = var_794, x = linear_22_cast_fp16)[name = tensor<string, []>("op_795_cast_fp16")];
            tensor<fp16, [256, 768]> layers_4_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64941184)))];
            tensor<fp16, [1, 256, 256]> linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_4_cross_attn_kv_proj_weight_to_fp16, x = memory_9_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<int32, [5]> var_799 = const()[name = tensor<string, []>("op_799"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_9_cast_fp16 = reshape(shape = var_799, x = linear_23_cast_fp16)[name = tensor<string, []>("kv_9_cast_fp16")];
            tensor<int32, [5]> var_803_begin_0 = const()[name = tensor<string, []>("op_803_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_803_end_0 = const()[name = tensor<string, []>("op_803_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_803_end_mask_0 = const()[name = tensor<string, []>("op_803_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_803_squeeze_mask_0 = const()[name = tensor<string, []>("op_803_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_803_cast_fp16 = slice_by_index(begin = var_803_begin_0, end = var_803_end_0, end_mask = var_803_end_mask_0, squeeze_mask = var_803_squeeze_mask_0, x = kv_9_cast_fp16)[name = tensor<string, []>("op_803_cast_fp16")];
            tensor<int32, [5]> var_807_begin_0 = const()[name = tensor<string, []>("op_807_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_807_end_0 = const()[name = tensor<string, []>("op_807_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_807_end_mask_0 = const()[name = tensor<string, []>("op_807_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_807_squeeze_mask_0 = const()[name = tensor<string, []>("op_807_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_807_cast_fp16 = slice_by_index(begin = var_807_begin_0, end = var_807_end_0, end_mask = var_807_end_mask_0, squeeze_mask = var_807_squeeze_mask_0, x = kv_9_cast_fp16)[name = tensor<string, []>("op_807_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = tensor<string, []>("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_810_transpose_x_0 = const()[name = tensor<string, []>("op_810_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_810_transpose_y_0 = const()[name = tensor<string, []>("op_810_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_85_perm_0 = const()[name = tensor<string, []>("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_803_cast_fp16)[name = tensor<string, []>("transpose_173")];
            tensor<fp16, [1, 1, 110, 128]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_795_cast_fp16)[name = tensor<string, []>("transpose_174")];
            tensor<fp16, [1, 1, 110, 256]> var_810_cast_fp16 = matmul(transpose_x = var_810_transpose_x_0, transpose_y = var_810_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor<string, []>("op_810_cast_fp16")];
            tensor<fp16, []> var_811_to_fp16 = const()[name = tensor<string, []>("op_811_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_37_cast_fp16 = mul(x = var_810_cast_fp16, y = var_811_to_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
            tensor<fp16, []> var_709_to_fp16 = const()[name = tensor<string, []>("op_709_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_91_cast_fp16 = select(a = var_709_to_fp16, b = attn_37_cast_fp16, cond = var_151)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_39_cast_fp16 = softmax(axis = var_718, x = input_91_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
            tensor<bool, []> out_19_transpose_x_0 = const()[name = tensor<string, []>("out_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_19_transpose_y_0 = const()[name = tensor<string, []>("out_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_807_cast_fp16)[name = tensor<string, []>("transpose_175")];
            tensor<fp16, [1, 1, 110, 128]> out_19_cast_fp16 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_39_cast_fp16, y = v_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<int32, [4]> var_819_perm_0 = const()[name = tensor<string, []>("op_819_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_820 = const()[name = tensor<string, []>("op_820"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_819_cast_fp16 = transpose(perm = var_819_perm_0, x = out_19_cast_fp16)[name = tensor<string, []>("transpose_172")];
            tensor<fp16, [1, 110, 128]> input_93_cast_fp16 = reshape(shape = var_820, x = var_819_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<fp16, [768, 128]> layers_4_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65334464)))];
            tensor<fp16, [1, 110, 768]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_cross_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_95_cast_fp16 = add(x = input_89_cast_fp16, y = linear_24_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<int32, [1]> x_45_axes_0 = const()[name = tensor<string, []>("x_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_4_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65531136)))];
            tensor<fp16, [1, 110, 768]> x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, epsilon = var_725_to_fp16, gamma = layers_4_norm_ff_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
            tensor<int32, [3]> input_97_perm_0 = const()[name = tensor<string, []>("input_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_99_pad_type_0 = const()[name = tensor<string, []>("input_99_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_99_strides_0 = const()[name = tensor<string, []>("input_99_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_99_pad_0 = const()[name = tensor<string, []>("input_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_99_dilations_0 = const()[name = tensor<string, []>("input_99_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_99_groups_0 = const()[name = tensor<string, []>("input_99_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_4_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65532736)))];
            tensor<fp16, [1, 768, 110]> input_97_cast_fp16 = transpose(perm = input_97_perm_0, x = x_45_cast_fp16)[name = tensor<string, []>("transpose_171")];
            tensor<fp16, [1, 3072, 110]> input_99_cast_fp16 = conv(dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = layers_4_ffn_conv1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<string, []> input_101_mode_0 = const()[name = tensor<string, []>("input_101_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_4_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70251392)))];
            tensor<fp16, [1, 768, 110]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = layers_4_ffn_conv2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
            tensor<int32, [3]> x_49_perm_0 = const()[name = tensor<string, []>("x_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_170")];
            tensor<fp16, [1, 110, 768]> input_103_cast_fp16 = add(x = input_95_cast_fp16, y = x_49_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<int32, [8]> k_padded_9_pad_0 = const()[name = tensor<string, []>("k_padded_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_9_mode_0 = const()[name = tensor<string, []>("k_padded_9_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_35_to_fp16 = const()[name = tensor<string, []>("const_35_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_9_cast_fp16 = pad(constant_val = const_35_to_fp16, mode = k_padded_9_mode_0, pad = k_padded_9_pad_0, x = k_17_cast_fp16)[name = tensor<string, []>("k_padded_9_cast_fp16")];
            tensor<int32, [8]> v_padded_9_pad_0 = const()[name = tensor<string, []>("v_padded_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_9_mode_0 = const()[name = tensor<string, []>("v_padded_9_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_36_to_fp16 = const()[name = tensor<string, []>("const_36_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_9_cast_fp16 = pad(constant_val = const_36_to_fp16, mode = v_padded_9_mode_0, pad = v_padded_9_pad_0, x = v_17_cast_fp16)[name = tensor<string, []>("v_padded_9_cast_fp16")];
            tensor<int32, []> var_872_axis_0 = const()[name = tensor<string, []>("op_872_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_872 = stack(axis = var_872_axis_0, values = (k_padded_9_cast_fp16, v_padded_9_cast_fp16))[name = tensor<string, []>("op_872_cast_fp16")];
            tensor<int32, []> var_884 = const()[name = tensor<string, []>("op_884"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_51_axes_0 = const()[name = tensor<string, []>("x_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_5_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74970048)))];
            tensor<fp16, []> var_891_to_fp16 = const()[name = tensor<string, []>("op_891_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_51_cast_fp16 = layer_norm(axes = x_51_axes_0, epsilon = var_891_to_fp16, gamma = layers_5_norm_sa_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_5_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74971648)))];
            tensor<fp16, [1, 110, 2304]> linear_25_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_qkv_proj_weight_to_fp16, x = x_51_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<int32, [5]> var_909 = const()[name = tensor<string, []>("op_909"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_11_cast_fp16 = reshape(shape = var_909, x = linear_25_cast_fp16)[name = tensor<string, []>("qkv_11_cast_fp16")];
            tensor<int32, [5]> q_31_begin_0 = const()[name = tensor<string, []>("q_31_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_31_end_0 = const()[name = tensor<string, []>("q_31_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_31_end_mask_0 = const()[name = tensor<string, []>("q_31_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_31_squeeze_mask_0 = const()[name = tensor<string, []>("q_31_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
            tensor<int32, [5]> k_21_begin_0 = const()[name = tensor<string, []>("k_21_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_21_end_0 = const()[name = tensor<string, []>("k_21_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_21_end_mask_0 = const()[name = tensor<string, []>("k_21_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_21_squeeze_mask_0 = const()[name = tensor<string, []>("k_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
            tensor<int32, [5]> v_21_begin_0 = const()[name = tensor<string, []>("v_21_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_21_end_0 = const()[name = tensor<string, []>("v_21_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_21_end_mask_0 = const()[name = tensor<string, []>("v_21_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_21_squeeze_mask_0 = const()[name = tensor<string, []>("v_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
            tensor<int32, [4]> v_t_11_perm_0 = const()[name = tensor<string, []>("v_t_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_924_transpose_x_0 = const()[name = tensor<string, []>("op_924_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_924_transpose_y_0 = const()[name = tensor<string, []>("op_924_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = tensor<string, []>("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_87_perm_0 = const()[name = tensor<string, []>("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_87 = transpose(perm = transpose_87_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_167")];
            tensor<fp16, [1, 12, 110, 64]> transpose_86 = transpose(perm = transpose_86_perm_0, x = q_31_cast_fp16)[name = tensor<string, []>("transpose_168")];
            tensor<fp16, [1, 12, 110, 110]> var_924_cast_fp16 = matmul(transpose_x = var_924_transpose_x_0, transpose_y = var_924_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor<string, []>("op_924_cast_fp16")];
            tensor<fp16, []> var_925_to_fp16 = const()[name = tensor<string, []>("op_925_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_41_cast_fp16 = mul(x = var_924_cast_fp16, y = var_925_to_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
            tensor<string, []> attn_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_41_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_41_cast_fp16_to_fp32 = cast(dtype = attn_41_cast_fp16_to_fp32_dtype_0, x = attn_41_cast_fp16)[name = tensor<string, []>("cast_83")];
            tensor<fp32, [1, 12, 110, 110]> input_105 = add(x = attn_41_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_105")];
            tensor<string, []> input_105_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_105_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_105_to_fp16 = cast(dtype = input_105_to_fp16_dtype_0, x = input_105)[name = tensor<string, []>("cast_82")];
            tensor<fp16, [1, 12, 110, 110]> attn_43_cast_fp16 = softmax(axis = var_884, x = input_105_to_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
            tensor<bool, []> out_21_transpose_x_0 = const()[name = tensor<string, []>("out_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_21_transpose_y_0 = const()[name = tensor<string, []>("out_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_11_cast_fp16 = transpose(perm = v_t_11_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_169")];
            tensor<fp16, [1, 12, 110, 64]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_43_cast_fp16, y = v_t_11_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<int32, [4]> var_936_perm_0 = const()[name = tensor<string, []>("op_936_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_937 = const()[name = tensor<string, []>("op_937"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_936_cast_fp16 = transpose(perm = var_936_perm_0, x = out_21_cast_fp16)[name = tensor<string, []>("transpose_166")];
            tensor<fp16, [1, 110, 768]> input_107_cast_fp16 = reshape(shape = var_937, x = var_936_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<fp16, [768, 768]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78510656)))];
            tensor<fp16, [1, 110, 768]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_26_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<int32, [1]> x_53_axes_0 = const()[name = tensor<string, []>("x_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_5_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79690368)))];
            tensor<fp16, [1, 110, 768]> x_53_cast_fp16 = layer_norm(axes = x_53_axes_0, epsilon = var_891_to_fp16, gamma = layers_5_norm_xa_query_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
            tensor<int32, [1]> memory_11_axes_0 = const()[name = tensor<string, []>("memory_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_5_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79691968)))];
            tensor<fp16, [1, 256, 768]> memory_11_cast_fp16 = layer_norm(axes = memory_11_axes_0, epsilon = var_891_to_fp16, gamma = layers_5_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_11_cast_fp16")];
            tensor<fp16, [128, 768]> layers_5_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79693568)))];
            tensor<fp16, [1, 110, 128]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_cross_attn_q_proj_weight_to_fp16, x = x_53_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<int32, [4]> var_960 = const()[name = tensor<string, []>("op_960"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_961_cast_fp16 = reshape(shape = var_960, x = linear_27_cast_fp16)[name = tensor<string, []>("op_961_cast_fp16")];
            tensor<fp16, [256, 768]> layers_5_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79890240)))];
            tensor<fp16, [1, 256, 256]> linear_28_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_5_cross_attn_kv_proj_weight_to_fp16, x = memory_11_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<int32, [5]> var_965 = const()[name = tensor<string, []>("op_965"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_11_cast_fp16 = reshape(shape = var_965, x = linear_28_cast_fp16)[name = tensor<string, []>("kv_11_cast_fp16")];
            tensor<int32, [5]> var_969_begin_0 = const()[name = tensor<string, []>("op_969_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_969_end_0 = const()[name = tensor<string, []>("op_969_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_969_end_mask_0 = const()[name = tensor<string, []>("op_969_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_969_squeeze_mask_0 = const()[name = tensor<string, []>("op_969_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_969_cast_fp16 = slice_by_index(begin = var_969_begin_0, end = var_969_end_0, end_mask = var_969_end_mask_0, squeeze_mask = var_969_squeeze_mask_0, x = kv_11_cast_fp16)[name = tensor<string, []>("op_969_cast_fp16")];
            tensor<int32, [5]> var_973_begin_0 = const()[name = tensor<string, []>("op_973_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_973_end_0 = const()[name = tensor<string, []>("op_973_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_973_end_mask_0 = const()[name = tensor<string, []>("op_973_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_973_squeeze_mask_0 = const()[name = tensor<string, []>("op_973_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_973_cast_fp16 = slice_by_index(begin = var_973_begin_0, end = var_973_end_0, end_mask = var_973_end_mask_0, squeeze_mask = var_973_squeeze_mask_0, x = kv_11_cast_fp16)[name = tensor<string, []>("op_973_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = tensor<string, []>("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_976_transpose_x_0 = const()[name = tensor<string, []>("op_976_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_976_transpose_y_0 = const()[name = tensor<string, []>("op_976_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = tensor<string, []>("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = tensor<string, []>("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_969_cast_fp16)[name = tensor<string, []>("transpose_163")];
            tensor<fp16, [1, 1, 110, 128]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_961_cast_fp16)[name = tensor<string, []>("transpose_164")];
            tensor<fp16, [1, 1, 110, 256]> var_976_cast_fp16 = matmul(transpose_x = var_976_transpose_x_0, transpose_y = var_976_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor<string, []>("op_976_cast_fp16")];
            tensor<fp16, []> var_977_to_fp16 = const()[name = tensor<string, []>("op_977_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_45_cast_fp16 = mul(x = var_976_cast_fp16, y = var_977_to_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
            tensor<fp16, []> var_875_to_fp16 = const()[name = tensor<string, []>("op_875_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_111_cast_fp16 = select(a = var_875_to_fp16, b = attn_45_cast_fp16, cond = var_151)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_47_cast_fp16 = softmax(axis = var_884, x = input_111_cast_fp16)[name = tensor<string, []>("attn_47_cast_fp16")];
            tensor<bool, []> out_23_transpose_x_0 = const()[name = tensor<string, []>("out_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_23_transpose_y_0 = const()[name = tensor<string, []>("out_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_973_cast_fp16)[name = tensor<string, []>("transpose_165")];
            tensor<fp16, [1, 1, 110, 128]> out_23_cast_fp16 = matmul(transpose_x = out_23_transpose_x_0, transpose_y = out_23_transpose_y_0, x = attn_47_cast_fp16, y = v_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<int32, [4]> var_985_perm_0 = const()[name = tensor<string, []>("op_985_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_986 = const()[name = tensor<string, []>("op_986"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_985_cast_fp16 = transpose(perm = var_985_perm_0, x = out_23_cast_fp16)[name = tensor<string, []>("transpose_162")];
            tensor<fp16, [1, 110, 128]> input_113_cast_fp16 = reshape(shape = var_986, x = var_985_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<fp16, [768, 128]> layers_5_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80283520)))];
            tensor<fp16, [1, 110, 768]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_cross_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_29_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<int32, [1]> x_55_axes_0 = const()[name = tensor<string, []>("x_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_5_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80480192)))];
            tensor<fp16, [1, 110, 768]> x_55_cast_fp16 = layer_norm(axes = x_55_axes_0, epsilon = var_891_to_fp16, gamma = layers_5_norm_ff_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")];
            tensor<int32, [3]> input_117_perm_0 = const()[name = tensor<string, []>("input_117_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_119_pad_type_0 = const()[name = tensor<string, []>("input_119_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_119_strides_0 = const()[name = tensor<string, []>("input_119_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_119_pad_0 = const()[name = tensor<string, []>("input_119_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_119_dilations_0 = const()[name = tensor<string, []>("input_119_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_119_groups_0 = const()[name = tensor<string, []>("input_119_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_5_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80481792)))];
            tensor<fp16, [1, 768, 110]> input_117_cast_fp16 = transpose(perm = input_117_perm_0, x = x_55_cast_fp16)[name = tensor<string, []>("transpose_161")];
            tensor<fp16, [1, 3072, 110]> input_119_cast_fp16 = conv(dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = layers_5_ffn_conv1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<string, []> input_121_mode_0 = const()[name = tensor<string, []>("input_121_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = input_119_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<string, []> x_57_pad_type_0 = const()[name = tensor<string, []>("x_57_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_57_strides_0 = const()[name = tensor<string, []>("x_57_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_57_pad_0 = const()[name = tensor<string, []>("x_57_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_57_dilations_0 = const()[name = tensor<string, []>("x_57_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_57_groups_0 = const()[name = tensor<string, []>("x_57_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_5_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85200448)))];
            tensor<fp16, [1, 768, 110]> x_57_cast_fp16 = conv(dilations = x_57_dilations_0, groups = x_57_groups_0, pad = x_57_pad_0, pad_type = x_57_pad_type_0, strides = x_57_strides_0, weight = layers_5_ffn_conv2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("x_57_cast_fp16")];
            tensor<int32, [3]> x_59_perm_0 = const()[name = tensor<string, []>("x_59_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = x_57_cast_fp16)[name = tensor<string, []>("transpose_160")];
            tensor<fp16, [1, 110, 768]> input_123_cast_fp16 = add(x = input_115_cast_fp16, y = x_59_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<int32, [8]> k_padded_11_pad_0 = const()[name = tensor<string, []>("k_padded_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_11_mode_0 = const()[name = tensor<string, []>("k_padded_11_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_42_to_fp16 = const()[name = tensor<string, []>("const_42_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_11_cast_fp16 = pad(constant_val = const_42_to_fp16, mode = k_padded_11_mode_0, pad = k_padded_11_pad_0, x = k_21_cast_fp16)[name = tensor<string, []>("k_padded_11_cast_fp16")];
            tensor<int32, [8]> v_padded_11_pad_0 = const()[name = tensor<string, []>("v_padded_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_11_mode_0 = const()[name = tensor<string, []>("v_padded_11_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_43_to_fp16 = const()[name = tensor<string, []>("const_43_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_11_cast_fp16 = pad(constant_val = const_43_to_fp16, mode = v_padded_11_mode_0, pad = v_padded_11_pad_0, x = v_21_cast_fp16)[name = tensor<string, []>("v_padded_11_cast_fp16")];
            tensor<int32, []> var_1038_axis_0 = const()[name = tensor<string, []>("op_1038_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1038 = stack(axis = var_1038_axis_0, values = (k_padded_11_cast_fp16, v_padded_11_cast_fp16))[name = tensor<string, []>("op_1038_cast_fp16")];
            tensor<int32, []> var_1050 = const()[name = tensor<string, []>("op_1050"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_61_axes_0 = const()[name = tensor<string, []>("x_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_6_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89919104)))];
            tensor<fp16, []> var_1057_to_fp16 = const()[name = tensor<string, []>("op_1057_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_61_cast_fp16 = layer_norm(axes = x_61_axes_0, epsilon = var_1057_to_fp16, gamma = layers_6_norm_sa_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_6_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89920704)))];
            tensor<fp16, [1, 110, 2304]> linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_qkv_proj_weight_to_fp16, x = x_61_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<int32, [5]> var_1075 = const()[name = tensor<string, []>("op_1075"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_13_cast_fp16 = reshape(shape = var_1075, x = linear_30_cast_fp16)[name = tensor<string, []>("qkv_13_cast_fp16")];
            tensor<int32, [5]> q_37_begin_0 = const()[name = tensor<string, []>("q_37_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_37_end_0 = const()[name = tensor<string, []>("q_37_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_37_end_mask_0 = const()[name = tensor<string, []>("q_37_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_37_squeeze_mask_0 = const()[name = tensor<string, []>("q_37_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
            tensor<int32, [5]> k_25_begin_0 = const()[name = tensor<string, []>("k_25_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_25_end_0 = const()[name = tensor<string, []>("k_25_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_25_end_mask_0 = const()[name = tensor<string, []>("k_25_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_25_squeeze_mask_0 = const()[name = tensor<string, []>("k_25_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
            tensor<int32, [5]> v_25_begin_0 = const()[name = tensor<string, []>("v_25_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_25_end_0 = const()[name = tensor<string, []>("v_25_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_25_end_mask_0 = const()[name = tensor<string, []>("v_25_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_25_squeeze_mask_0 = const()[name = tensor<string, []>("v_25_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
            tensor<int32, [4]> v_t_13_perm_0 = const()[name = tensor<string, []>("v_t_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1090_transpose_x_0 = const()[name = tensor<string, []>("op_1090_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1090_transpose_y_0 = const()[name = tensor<string, []>("op_1090_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = tensor<string, []>("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_91_perm_0 = const()[name = tensor<string, []>("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_91 = transpose(perm = transpose_91_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_157")];
            tensor<fp16, [1, 12, 110, 64]> transpose_90 = transpose(perm = transpose_90_perm_0, x = q_37_cast_fp16)[name = tensor<string, []>("transpose_158")];
            tensor<fp16, [1, 12, 110, 110]> var_1090_cast_fp16 = matmul(transpose_x = var_1090_transpose_x_0, transpose_y = var_1090_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor<string, []>("op_1090_cast_fp16")];
            tensor<fp16, []> var_1091_to_fp16 = const()[name = tensor<string, []>("op_1091_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_49_cast_fp16 = mul(x = var_1090_cast_fp16, y = var_1091_to_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
            tensor<string, []> attn_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_49_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_49_cast_fp16_to_fp32 = cast(dtype = attn_49_cast_fp16_to_fp32_dtype_0, x = attn_49_cast_fp16)[name = tensor<string, []>("cast_81")];
            tensor<fp32, [1, 12, 110, 110]> input_125 = add(x = attn_49_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_125")];
            tensor<string, []> input_125_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_125_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_125_to_fp16 = cast(dtype = input_125_to_fp16_dtype_0, x = input_125)[name = tensor<string, []>("cast_80")];
            tensor<fp16, [1, 12, 110, 110]> attn_51_cast_fp16 = softmax(axis = var_1050, x = input_125_to_fp16)[name = tensor<string, []>("attn_51_cast_fp16")];
            tensor<bool, []> out_25_transpose_x_0 = const()[name = tensor<string, []>("out_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_25_transpose_y_0 = const()[name = tensor<string, []>("out_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_13_cast_fp16 = transpose(perm = v_t_13_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_159")];
            tensor<fp16, [1, 12, 110, 64]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = attn_51_cast_fp16, y = v_t_13_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<int32, [4]> var_1102_perm_0 = const()[name = tensor<string, []>("op_1102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1103 = const()[name = tensor<string, []>("op_1103"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_1102_cast_fp16 = transpose(perm = var_1102_perm_0, x = out_25_cast_fp16)[name = tensor<string, []>("transpose_156")];
            tensor<fp16, [1, 110, 768]> input_127_cast_fp16 = reshape(shape = var_1103, x = var_1102_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<fp16, [768, 768]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93459712)))];
            tensor<fp16, [1, 110, 768]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_31_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [1]> x_63_axes_0 = const()[name = tensor<string, []>("x_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_6_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94639424)))];
            tensor<fp16, [1, 110, 768]> x_63_cast_fp16 = layer_norm(axes = x_63_axes_0, epsilon = var_1057_to_fp16, gamma = layers_6_norm_xa_query_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<int32, [1]> memory_13_axes_0 = const()[name = tensor<string, []>("memory_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_6_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94641024)))];
            tensor<fp16, [1, 256, 768]> memory_13_cast_fp16 = layer_norm(axes = memory_13_axes_0, epsilon = var_1057_to_fp16, gamma = layers_6_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_13_cast_fp16")];
            tensor<fp16, [128, 768]> layers_6_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94642624)))];
            tensor<fp16, [1, 110, 128]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_cross_attn_q_proj_weight_to_fp16, x = x_63_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [4]> var_1126 = const()[name = tensor<string, []>("op_1126"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_1127_cast_fp16 = reshape(shape = var_1126, x = linear_32_cast_fp16)[name = tensor<string, []>("op_1127_cast_fp16")];
            tensor<fp16, [256, 768]> layers_6_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94839296)))];
            tensor<fp16, [1, 256, 256]> linear_33_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_6_cross_attn_kv_proj_weight_to_fp16, x = memory_13_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<int32, [5]> var_1131 = const()[name = tensor<string, []>("op_1131"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_13_cast_fp16 = reshape(shape = var_1131, x = linear_33_cast_fp16)[name = tensor<string, []>("kv_13_cast_fp16")];
            tensor<int32, [5]> var_1135_begin_0 = const()[name = tensor<string, []>("op_1135_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1135_end_0 = const()[name = tensor<string, []>("op_1135_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_1135_end_mask_0 = const()[name = tensor<string, []>("op_1135_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1135_squeeze_mask_0 = const()[name = tensor<string, []>("op_1135_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1135_cast_fp16 = slice_by_index(begin = var_1135_begin_0, end = var_1135_end_0, end_mask = var_1135_end_mask_0, squeeze_mask = var_1135_squeeze_mask_0, x = kv_13_cast_fp16)[name = tensor<string, []>("op_1135_cast_fp16")];
            tensor<int32, [5]> var_1139_begin_0 = const()[name = tensor<string, []>("op_1139_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_1139_end_0 = const()[name = tensor<string, []>("op_1139_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_1139_end_mask_0 = const()[name = tensor<string, []>("op_1139_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1139_squeeze_mask_0 = const()[name = tensor<string, []>("op_1139_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1139_cast_fp16 = slice_by_index(begin = var_1139_begin_0, end = var_1139_end_0, end_mask = var_1139_end_mask_0, squeeze_mask = var_1139_squeeze_mask_0, x = kv_13_cast_fp16)[name = tensor<string, []>("op_1139_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = tensor<string, []>("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1142_transpose_x_0 = const()[name = tensor<string, []>("op_1142_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1142_transpose_y_0 = const()[name = tensor<string, []>("op_1142_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = tensor<string, []>("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_93_perm_0 = const()[name = tensor<string, []>("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_1135_cast_fp16)[name = tensor<string, []>("transpose_153")];
            tensor<fp16, [1, 1, 110, 128]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_1127_cast_fp16)[name = tensor<string, []>("transpose_154")];
            tensor<fp16, [1, 1, 110, 256]> var_1142_cast_fp16 = matmul(transpose_x = var_1142_transpose_x_0, transpose_y = var_1142_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor<string, []>("op_1142_cast_fp16")];
            tensor<fp16, []> var_1143_to_fp16 = const()[name = tensor<string, []>("op_1143_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_53_cast_fp16 = mul(x = var_1142_cast_fp16, y = var_1143_to_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
            tensor<fp16, []> var_1041_to_fp16 = const()[name = tensor<string, []>("op_1041_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_131_cast_fp16 = select(a = var_1041_to_fp16, b = attn_53_cast_fp16, cond = var_151)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_55_cast_fp16 = softmax(axis = var_1050, x = input_131_cast_fp16)[name = tensor<string, []>("attn_55_cast_fp16")];
            tensor<bool, []> out_27_transpose_x_0 = const()[name = tensor<string, []>("out_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_27_transpose_y_0 = const()[name = tensor<string, []>("out_27_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_1139_cast_fp16)[name = tensor<string, []>("transpose_155")];
            tensor<fp16, [1, 1, 110, 128]> out_27_cast_fp16 = matmul(transpose_x = out_27_transpose_x_0, transpose_y = out_27_transpose_y_0, x = attn_55_cast_fp16, y = v_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<int32, [4]> var_1151_perm_0 = const()[name = tensor<string, []>("op_1151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1152 = const()[name = tensor<string, []>("op_1152"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_1151_cast_fp16 = transpose(perm = var_1151_perm_0, x = out_27_cast_fp16)[name = tensor<string, []>("transpose_152")];
            tensor<fp16, [1, 110, 128]> input_133_cast_fp16 = reshape(shape = var_1152, x = var_1151_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<fp16, [768, 128]> layers_6_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95232576)))];
            tensor<fp16, [1, 110, 768]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_cross_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<int32, [1]> x_65_axes_0 = const()[name = tensor<string, []>("x_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_6_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95429248)))];
            tensor<fp16, [1, 110, 768]> x_65_cast_fp16 = layer_norm(axes = x_65_axes_0, epsilon = var_1057_to_fp16, gamma = layers_6_norm_ff_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<int32, [3]> input_137_perm_0 = const()[name = tensor<string, []>("input_137_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_139_pad_type_0 = const()[name = tensor<string, []>("input_139_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_139_strides_0 = const()[name = tensor<string, []>("input_139_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_139_pad_0 = const()[name = tensor<string, []>("input_139_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_139_dilations_0 = const()[name = tensor<string, []>("input_139_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_139_groups_0 = const()[name = tensor<string, []>("input_139_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_6_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95430848)))];
            tensor<fp16, [1, 768, 110]> input_137_cast_fp16 = transpose(perm = input_137_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_151")];
            tensor<fp16, [1, 3072, 110]> input_139_cast_fp16 = conv(dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = layers_6_ffn_conv1_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<string, []> input_141_mode_0 = const()[name = tensor<string, []>("input_141_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = input_139_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<string, []> x_67_pad_type_0 = const()[name = tensor<string, []>("x_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_67_strides_0 = const()[name = tensor<string, []>("x_67_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_67_pad_0 = const()[name = tensor<string, []>("x_67_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_67_dilations_0 = const()[name = tensor<string, []>("x_67_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_67_groups_0 = const()[name = tensor<string, []>("x_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_6_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100149504)))];
            tensor<fp16, [1, 768, 110]> x_67_cast_fp16 = conv(dilations = x_67_dilations_0, groups = x_67_groups_0, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = x_67_strides_0, weight = layers_6_ffn_conv2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<int32, [3]> x_69_perm_0 = const()[name = tensor<string, []>("x_69_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_69_cast_fp16 = transpose(perm = x_69_perm_0, x = x_67_cast_fp16)[name = tensor<string, []>("transpose_150")];
            tensor<fp16, [1, 110, 768]> input_143_cast_fp16 = add(x = input_135_cast_fp16, y = x_69_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [8]> k_padded_13_pad_0 = const()[name = tensor<string, []>("k_padded_13_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_13_mode_0 = const()[name = tensor<string, []>("k_padded_13_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_49_to_fp16 = const()[name = tensor<string, []>("const_49_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_13_cast_fp16 = pad(constant_val = const_49_to_fp16, mode = k_padded_13_mode_0, pad = k_padded_13_pad_0, x = k_25_cast_fp16)[name = tensor<string, []>("k_padded_13_cast_fp16")];
            tensor<int32, [8]> v_padded_13_pad_0 = const()[name = tensor<string, []>("v_padded_13_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_13_mode_0 = const()[name = tensor<string, []>("v_padded_13_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_50_to_fp16 = const()[name = tensor<string, []>("const_50_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_13_cast_fp16 = pad(constant_val = const_50_to_fp16, mode = v_padded_13_mode_0, pad = v_padded_13_pad_0, x = v_25_cast_fp16)[name = tensor<string, []>("v_padded_13_cast_fp16")];
            tensor<int32, []> var_1204_axis_0 = const()[name = tensor<string, []>("op_1204_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1204 = stack(axis = var_1204_axis_0, values = (k_padded_13_cast_fp16, v_padded_13_cast_fp16))[name = tensor<string, []>("op_1204_cast_fp16")];
            tensor<int32, []> var_1216 = const()[name = tensor<string, []>("op_1216"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_71_axes_0 = const()[name = tensor<string, []>("x_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_7_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104868160)))];
            tensor<fp16, []> var_1223_to_fp16 = const()[name = tensor<string, []>("op_1223_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_71_cast_fp16 = layer_norm(axes = x_71_axes_0, epsilon = var_1223_to_fp16, gamma = layers_7_norm_sa_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_7_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104869760)))];
            tensor<fp16, [1, 110, 2304]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_qkv_proj_weight_to_fp16, x = x_71_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<int32, [5]> var_1241 = const()[name = tensor<string, []>("op_1241"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_15_cast_fp16 = reshape(shape = var_1241, x = linear_35_cast_fp16)[name = tensor<string, []>("qkv_15_cast_fp16")];
            tensor<int32, [5]> q_43_begin_0 = const()[name = tensor<string, []>("q_43_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_43_end_0 = const()[name = tensor<string, []>("q_43_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_43_end_mask_0 = const()[name = tensor<string, []>("q_43_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_43_squeeze_mask_0 = const()[name = tensor<string, []>("q_43_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
            tensor<int32, [5]> k_29_begin_0 = const()[name = tensor<string, []>("k_29_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_29_end_0 = const()[name = tensor<string, []>("k_29_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_29_end_mask_0 = const()[name = tensor<string, []>("k_29_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_29_squeeze_mask_0 = const()[name = tensor<string, []>("k_29_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
            tensor<int32, [5]> v_29_begin_0 = const()[name = tensor<string, []>("v_29_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_29_end_0 = const()[name = tensor<string, []>("v_29_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_29_end_mask_0 = const()[name = tensor<string, []>("v_29_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_29_squeeze_mask_0 = const()[name = tensor<string, []>("v_29_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
            tensor<int32, [4]> v_t_15_perm_0 = const()[name = tensor<string, []>("v_t_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1256_transpose_x_0 = const()[name = tensor<string, []>("op_1256_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1256_transpose_y_0 = const()[name = tensor<string, []>("op_1256_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = tensor<string, []>("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_95_perm_0 = const()[name = tensor<string, []>("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_95 = transpose(perm = transpose_95_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_147")];
            tensor<fp16, [1, 12, 110, 64]> transpose_94 = transpose(perm = transpose_94_perm_0, x = q_43_cast_fp16)[name = tensor<string, []>("transpose_148")];
            tensor<fp16, [1, 12, 110, 110]> var_1256_cast_fp16 = matmul(transpose_x = var_1256_transpose_x_0, transpose_y = var_1256_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor<string, []>("op_1256_cast_fp16")];
            tensor<fp16, []> var_1257_to_fp16 = const()[name = tensor<string, []>("op_1257_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_57_cast_fp16 = mul(x = var_1256_cast_fp16, y = var_1257_to_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
            tensor<string, []> attn_57_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_57_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_57_cast_fp16_to_fp32 = cast(dtype = attn_57_cast_fp16_to_fp32_dtype_0, x = attn_57_cast_fp16)[name = tensor<string, []>("cast_79")];
            tensor<fp32, [1, 12, 110, 110]> input_145 = add(x = attn_57_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_145")];
            tensor<string, []> input_145_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_145_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_145_to_fp16 = cast(dtype = input_145_to_fp16_dtype_0, x = input_145)[name = tensor<string, []>("cast_78")];
            tensor<fp16, [1, 12, 110, 110]> attn_59_cast_fp16 = softmax(axis = var_1216, x = input_145_to_fp16)[name = tensor<string, []>("attn_59_cast_fp16")];
            tensor<bool, []> out_29_transpose_x_0 = const()[name = tensor<string, []>("out_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_29_transpose_y_0 = const()[name = tensor<string, []>("out_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_15_cast_fp16 = transpose(perm = v_t_15_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_149")];
            tensor<fp16, [1, 12, 110, 64]> out_29_cast_fp16 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = attn_59_cast_fp16, y = v_t_15_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<int32, [4]> var_1268_perm_0 = const()[name = tensor<string, []>("op_1268_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1269 = const()[name = tensor<string, []>("op_1269"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_1268_cast_fp16 = transpose(perm = var_1268_perm_0, x = out_29_cast_fp16)[name = tensor<string, []>("transpose_146")];
            tensor<fp16, [1, 110, 768]> input_147_cast_fp16 = reshape(shape = var_1269, x = var_1268_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<fp16, [768, 768]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108408768)))];
            tensor<fp16, [1, 110, 768]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_149_cast_fp16 = add(x = input_143_cast_fp16, y = linear_36_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<int32, [1]> x_73_axes_0 = const()[name = tensor<string, []>("x_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_7_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109588480)))];
            tensor<fp16, [1, 110, 768]> x_73_cast_fp16 = layer_norm(axes = x_73_axes_0, epsilon = var_1223_to_fp16, gamma = layers_7_norm_xa_query_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
            tensor<int32, [1]> memory_15_axes_0 = const()[name = tensor<string, []>("memory_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_7_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109590080)))];
            tensor<fp16, [1, 256, 768]> memory_15_cast_fp16 = layer_norm(axes = memory_15_axes_0, epsilon = var_1223_to_fp16, gamma = layers_7_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_15_cast_fp16")];
            tensor<fp16, [128, 768]> layers_7_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109591680)))];
            tensor<fp16, [1, 110, 128]> linear_37_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_cross_attn_q_proj_weight_to_fp16, x = x_73_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<int32, [4]> var_1292 = const()[name = tensor<string, []>("op_1292"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_1293_cast_fp16 = reshape(shape = var_1292, x = linear_37_cast_fp16)[name = tensor<string, []>("op_1293_cast_fp16")];
            tensor<fp16, [256, 768]> layers_7_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109788352)))];
            tensor<fp16, [1, 256, 256]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_7_cross_attn_kv_proj_weight_to_fp16, x = memory_15_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<int32, [5]> var_1297 = const()[name = tensor<string, []>("op_1297"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_15_cast_fp16 = reshape(shape = var_1297, x = linear_38_cast_fp16)[name = tensor<string, []>("kv_15_cast_fp16")];
            tensor<int32, [5]> var_1301_begin_0 = const()[name = tensor<string, []>("op_1301_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1301_end_0 = const()[name = tensor<string, []>("op_1301_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_1301_end_mask_0 = const()[name = tensor<string, []>("op_1301_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1301_squeeze_mask_0 = const()[name = tensor<string, []>("op_1301_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1301_cast_fp16 = slice_by_index(begin = var_1301_begin_0, end = var_1301_end_0, end_mask = var_1301_end_mask_0, squeeze_mask = var_1301_squeeze_mask_0, x = kv_15_cast_fp16)[name = tensor<string, []>("op_1301_cast_fp16")];
            tensor<int32, [5]> var_1305_begin_0 = const()[name = tensor<string, []>("op_1305_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_1305_end_0 = const()[name = tensor<string, []>("op_1305_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_1305_end_mask_0 = const()[name = tensor<string, []>("op_1305_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1305_squeeze_mask_0 = const()[name = tensor<string, []>("op_1305_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1305_cast_fp16 = slice_by_index(begin = var_1305_begin_0, end = var_1305_end_0, end_mask = var_1305_end_mask_0, squeeze_mask = var_1305_squeeze_mask_0, x = kv_15_cast_fp16)[name = tensor<string, []>("op_1305_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = tensor<string, []>("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1308_transpose_x_0 = const()[name = tensor<string, []>("op_1308_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1308_transpose_y_0 = const()[name = tensor<string, []>("op_1308_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_97 = transpose(perm = transpose_97_perm_0, x = var_1301_cast_fp16)[name = tensor<string, []>("transpose_143")];
            tensor<fp16, [1, 1, 110, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_1293_cast_fp16)[name = tensor<string, []>("transpose_144")];
            tensor<fp16, [1, 1, 110, 256]> var_1308_cast_fp16 = matmul(transpose_x = var_1308_transpose_x_0, transpose_y = var_1308_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor<string, []>("op_1308_cast_fp16")];
            tensor<fp16, []> var_1309_to_fp16 = const()[name = tensor<string, []>("op_1309_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_61_cast_fp16 = mul(x = var_1308_cast_fp16, y = var_1309_to_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
            tensor<fp16, []> var_1207_to_fp16 = const()[name = tensor<string, []>("op_1207_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_151_cast_fp16 = select(a = var_1207_to_fp16, b = attn_61_cast_fp16, cond = var_151)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_63_cast_fp16 = softmax(axis = var_1216, x = input_151_cast_fp16)[name = tensor<string, []>("attn_63_cast_fp16")];
            tensor<bool, []> out_31_transpose_x_0 = const()[name = tensor<string, []>("out_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_31_transpose_y_0 = const()[name = tensor<string, []>("out_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_1305_cast_fp16)[name = tensor<string, []>("transpose_145")];
            tensor<fp16, [1, 1, 110, 128]> out_31_cast_fp16 = matmul(transpose_x = out_31_transpose_x_0, transpose_y = out_31_transpose_y_0, x = attn_63_cast_fp16, y = v_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<int32, [4]> var_1317_perm_0 = const()[name = tensor<string, []>("op_1317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1318 = const()[name = tensor<string, []>("op_1318"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_1317_cast_fp16 = transpose(perm = var_1317_perm_0, x = out_31_cast_fp16)[name = tensor<string, []>("transpose_142")];
            tensor<fp16, [1, 110, 128]> input_153_cast_fp16 = reshape(shape = var_1318, x = var_1317_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<fp16, [768, 128]> layers_7_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110181632)))];
            tensor<fp16, [1, 110, 768]> linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_cross_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_155_cast_fp16 = add(x = input_149_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_7_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110378304)))];
            tensor<fp16, [1, 110, 768]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, epsilon = var_1223_to_fp16, gamma = layers_7_norm_ff_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<int32, [3]> input_157_perm_0 = const()[name = tensor<string, []>("input_157_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_159_pad_type_0 = const()[name = tensor<string, []>("input_159_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_159_strides_0 = const()[name = tensor<string, []>("input_159_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_159_pad_0 = const()[name = tensor<string, []>("input_159_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_159_dilations_0 = const()[name = tensor<string, []>("input_159_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_159_groups_0 = const()[name = tensor<string, []>("input_159_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_7_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110379904)))];
            tensor<fp16, [1, 768, 110]> input_157_cast_fp16 = transpose(perm = input_157_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_141")];
            tensor<fp16, [1, 3072, 110]> input_159_cast_fp16 = conv(dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = layers_7_ffn_conv1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<string, []> input_161_mode_0 = const()[name = tensor<string, []>("input_161_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = input_159_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_7_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115098560)))];
            tensor<fp16, [1, 768, 110]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = layers_7_ffn_conv2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<int32, [3]> x_79_perm_0 = const()[name = tensor<string, []>("x_79_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_140")];
            tensor<fp16, [1, 110, 768]> input_163_cast_fp16 = add(x = input_155_cast_fp16, y = x_79_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<int32, [8]> k_padded_15_pad_0 = const()[name = tensor<string, []>("k_padded_15_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_15_mode_0 = const()[name = tensor<string, []>("k_padded_15_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_56_to_fp16 = const()[name = tensor<string, []>("const_56_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_15_cast_fp16 = pad(constant_val = const_56_to_fp16, mode = k_padded_15_mode_0, pad = k_padded_15_pad_0, x = k_29_cast_fp16)[name = tensor<string, []>("k_padded_15_cast_fp16")];
            tensor<int32, [8]> v_padded_15_pad_0 = const()[name = tensor<string, []>("v_padded_15_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_15_mode_0 = const()[name = tensor<string, []>("v_padded_15_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_57_to_fp16 = const()[name = tensor<string, []>("const_57_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_15_cast_fp16 = pad(constant_val = const_57_to_fp16, mode = v_padded_15_mode_0, pad = v_padded_15_pad_0, x = v_29_cast_fp16)[name = tensor<string, []>("v_padded_15_cast_fp16")];
            tensor<int32, []> var_1370_axis_0 = const()[name = tensor<string, []>("op_1370_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1370 = stack(axis = var_1370_axis_0, values = (k_padded_15_cast_fp16, v_padded_15_cast_fp16))[name = tensor<string, []>("op_1370_cast_fp16")];
            tensor<int32, []> var_1382 = const()[name = tensor<string, []>("op_1382"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_81_axes_0 = const()[name = tensor<string, []>("x_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_8_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119817216)))];
            tensor<fp16, []> var_1389_to_fp16 = const()[name = tensor<string, []>("op_1389_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_81_cast_fp16 = layer_norm(axes = x_81_axes_0, epsilon = var_1389_to_fp16, gamma = layers_8_norm_sa_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_81_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_8_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119818816)))];
            tensor<fp16, [1, 110, 2304]> linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_qkv_proj_weight_to_fp16, x = x_81_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<int32, [5]> var_1407 = const()[name = tensor<string, []>("op_1407"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_17_cast_fp16 = reshape(shape = var_1407, x = linear_40_cast_fp16)[name = tensor<string, []>("qkv_17_cast_fp16")];
            tensor<int32, [5]> q_49_begin_0 = const()[name = tensor<string, []>("q_49_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_49_end_0 = const()[name = tensor<string, []>("q_49_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_49_end_mask_0 = const()[name = tensor<string, []>("q_49_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_49_squeeze_mask_0 = const()[name = tensor<string, []>("q_49_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
            tensor<int32, [5]> k_33_begin_0 = const()[name = tensor<string, []>("k_33_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_33_end_0 = const()[name = tensor<string, []>("k_33_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_33_end_mask_0 = const()[name = tensor<string, []>("k_33_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_33_squeeze_mask_0 = const()[name = tensor<string, []>("k_33_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
            tensor<int32, [5]> v_33_begin_0 = const()[name = tensor<string, []>("v_33_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_33_end_0 = const()[name = tensor<string, []>("v_33_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_33_end_mask_0 = const()[name = tensor<string, []>("v_33_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_33_squeeze_mask_0 = const()[name = tensor<string, []>("v_33_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor<string, []>("v_33_cast_fp16")];
            tensor<int32, [4]> v_t_17_perm_0 = const()[name = tensor<string, []>("v_t_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1422_transpose_x_0 = const()[name = tensor<string, []>("op_1422_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1422_transpose_y_0 = const()[name = tensor<string, []>("op_1422_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = tensor<string, []>("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_137")];
            tensor<fp16, [1, 12, 110, 64]> transpose_98 = transpose(perm = transpose_98_perm_0, x = q_49_cast_fp16)[name = tensor<string, []>("transpose_138")];
            tensor<fp16, [1, 12, 110, 110]> var_1422_cast_fp16 = matmul(transpose_x = var_1422_transpose_x_0, transpose_y = var_1422_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor<string, []>("op_1422_cast_fp16")];
            tensor<fp16, []> var_1423_to_fp16 = const()[name = tensor<string, []>("op_1423_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_65_cast_fp16 = mul(x = var_1422_cast_fp16, y = var_1423_to_fp16)[name = tensor<string, []>("attn_65_cast_fp16")];
            tensor<string, []> attn_65_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_65_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_65_cast_fp16_to_fp32 = cast(dtype = attn_65_cast_fp16_to_fp32_dtype_0, x = attn_65_cast_fp16)[name = tensor<string, []>("cast_77")];
            tensor<fp32, [1, 12, 110, 110]> input_165 = add(x = attn_65_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_165")];
            tensor<string, []> input_165_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_165_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_165_to_fp16 = cast(dtype = input_165_to_fp16_dtype_0, x = input_165)[name = tensor<string, []>("cast_76")];
            tensor<fp16, [1, 12, 110, 110]> attn_67_cast_fp16 = softmax(axis = var_1382, x = input_165_to_fp16)[name = tensor<string, []>("attn_67_cast_fp16")];
            tensor<bool, []> out_33_transpose_x_0 = const()[name = tensor<string, []>("out_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_33_transpose_y_0 = const()[name = tensor<string, []>("out_33_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_17_cast_fp16 = transpose(perm = v_t_17_perm_0, x = v_33_cast_fp16)[name = tensor<string, []>("transpose_139")];
            tensor<fp16, [1, 12, 110, 64]> out_33_cast_fp16 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = attn_67_cast_fp16, y = v_t_17_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<int32, [4]> var_1434_perm_0 = const()[name = tensor<string, []>("op_1434_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1435 = const()[name = tensor<string, []>("op_1435"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_1434_cast_fp16 = transpose(perm = var_1434_perm_0, x = out_33_cast_fp16)[name = tensor<string, []>("transpose_136")];
            tensor<fp16, [1, 110, 768]> input_167_cast_fp16 = reshape(shape = var_1435, x = var_1434_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<fp16, [768, 768]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123357824)))];
            tensor<fp16, [1, 110, 768]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_41_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<int32, [1]> x_83_axes_0 = const()[name = tensor<string, []>("x_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_8_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124537536)))];
            tensor<fp16, [1, 110, 768]> x_83_cast_fp16 = layer_norm(axes = x_83_axes_0, epsilon = var_1389_to_fp16, gamma = layers_8_norm_xa_query_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
            tensor<int32, [1]> memory_17_axes_0 = const()[name = tensor<string, []>("memory_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_8_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124539136)))];
            tensor<fp16, [1, 256, 768]> memory_17_cast_fp16 = layer_norm(axes = memory_17_axes_0, epsilon = var_1389_to_fp16, gamma = layers_8_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_17_cast_fp16")];
            tensor<fp16, [128, 768]> layers_8_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124540736)))];
            tensor<fp16, [1, 110, 128]> linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_8_cross_attn_q_proj_weight_to_fp16, x = x_83_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1458 = const()[name = tensor<string, []>("op_1458"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_1459_cast_fp16 = reshape(shape = var_1458, x = linear_42_cast_fp16)[name = tensor<string, []>("op_1459_cast_fp16")];
            tensor<fp16, [256, 768]> layers_8_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124737408)))];
            tensor<fp16, [1, 256, 256]> linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_8_cross_attn_kv_proj_weight_to_fp16, x = memory_17_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<int32, [5]> var_1463 = const()[name = tensor<string, []>("op_1463"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_17_cast_fp16 = reshape(shape = var_1463, x = linear_43_cast_fp16)[name = tensor<string, []>("kv_17_cast_fp16")];
            tensor<int32, [5]> var_1467_begin_0 = const()[name = tensor<string, []>("op_1467_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1467_end_0 = const()[name = tensor<string, []>("op_1467_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_1467_end_mask_0 = const()[name = tensor<string, []>("op_1467_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1467_squeeze_mask_0 = const()[name = tensor<string, []>("op_1467_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1467_cast_fp16 = slice_by_index(begin = var_1467_begin_0, end = var_1467_end_0, end_mask = var_1467_end_mask_0, squeeze_mask = var_1467_squeeze_mask_0, x = kv_17_cast_fp16)[name = tensor<string, []>("op_1467_cast_fp16")];
            tensor<int32, [5]> var_1471_begin_0 = const()[name = tensor<string, []>("op_1471_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_1471_end_0 = const()[name = tensor<string, []>("op_1471_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_1471_end_mask_0 = const()[name = tensor<string, []>("op_1471_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1471_squeeze_mask_0 = const()[name = tensor<string, []>("op_1471_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1471_cast_fp16 = slice_by_index(begin = var_1471_begin_0, end = var_1471_end_0, end_mask = var_1471_end_mask_0, squeeze_mask = var_1471_squeeze_mask_0, x = kv_17_cast_fp16)[name = tensor<string, []>("op_1471_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = tensor<string, []>("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1474_transpose_x_0 = const()[name = tensor<string, []>("op_1474_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1474_transpose_y_0 = const()[name = tensor<string, []>("op_1474_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = tensor<string, []>("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_101 = transpose(perm = transpose_101_perm_0, x = var_1467_cast_fp16)[name = tensor<string, []>("transpose_133")];
            tensor<fp16, [1, 1, 110, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_1459_cast_fp16)[name = tensor<string, []>("transpose_134")];
            tensor<fp16, [1, 1, 110, 256]> var_1474_cast_fp16 = matmul(transpose_x = var_1474_transpose_x_0, transpose_y = var_1474_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor<string, []>("op_1474_cast_fp16")];
            tensor<fp16, []> var_1475_to_fp16 = const()[name = tensor<string, []>("op_1475_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_69_cast_fp16 = mul(x = var_1474_cast_fp16, y = var_1475_to_fp16)[name = tensor<string, []>("attn_69_cast_fp16")];
            tensor<fp16, []> var_1373_to_fp16 = const()[name = tensor<string, []>("op_1373_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_171_cast_fp16 = select(a = var_1373_to_fp16, b = attn_69_cast_fp16, cond = var_151)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_71_cast_fp16 = softmax(axis = var_1382, x = input_171_cast_fp16)[name = tensor<string, []>("attn_71_cast_fp16")];
            tensor<bool, []> out_35_transpose_x_0 = const()[name = tensor<string, []>("out_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_35_transpose_y_0 = const()[name = tensor<string, []>("out_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_1471_cast_fp16)[name = tensor<string, []>("transpose_135")];
            tensor<fp16, [1, 1, 110, 128]> out_35_cast_fp16 = matmul(transpose_x = out_35_transpose_x_0, transpose_y = out_35_transpose_y_0, x = attn_71_cast_fp16, y = v_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<int32, [4]> var_1483_perm_0 = const()[name = tensor<string, []>("op_1483_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1484 = const()[name = tensor<string, []>("op_1484"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_1483_cast_fp16 = transpose(perm = var_1483_perm_0, x = out_35_cast_fp16)[name = tensor<string, []>("transpose_132")];
            tensor<fp16, [1, 110, 128]> input_173_cast_fp16 = reshape(shape = var_1484, x = var_1483_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<fp16, [768, 128]> layers_8_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125130688)))];
            tensor<fp16, [1, 110, 768]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_cross_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_44_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<int32, [1]> x_85_axes_0 = const()[name = tensor<string, []>("x_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_8_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125327360)))];
            tensor<fp16, [1, 110, 768]> x_85_cast_fp16 = layer_norm(axes = x_85_axes_0, epsilon = var_1389_to_fp16, gamma = layers_8_norm_ff_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<int32, [3]> input_177_perm_0 = const()[name = tensor<string, []>("input_177_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_179_pad_type_0 = const()[name = tensor<string, []>("input_179_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_179_strides_0 = const()[name = tensor<string, []>("input_179_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_179_pad_0 = const()[name = tensor<string, []>("input_179_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_179_dilations_0 = const()[name = tensor<string, []>("input_179_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_179_groups_0 = const()[name = tensor<string, []>("input_179_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_8_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125328960)))];
            tensor<fp16, [1, 768, 110]> input_177_cast_fp16 = transpose(perm = input_177_perm_0, x = x_85_cast_fp16)[name = tensor<string, []>("transpose_131")];
            tensor<fp16, [1, 3072, 110]> input_179_cast_fp16 = conv(dilations = input_179_dilations_0, groups = input_179_groups_0, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = input_179_strides_0, weight = layers_8_ffn_conv1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<string, []> input_181_mode_0 = const()[name = tensor<string, []>("input_181_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<string, []> x_87_pad_type_0 = const()[name = tensor<string, []>("x_87_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_87_strides_0 = const()[name = tensor<string, []>("x_87_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_87_dilations_0 = const()[name = tensor<string, []>("x_87_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_87_groups_0 = const()[name = tensor<string, []>("x_87_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_8_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130047616)))];
            tensor<fp16, [1, 768, 110]> x_87_cast_fp16 = conv(dilations = x_87_dilations_0, groups = x_87_groups_0, pad = x_87_pad_0, pad_type = x_87_pad_type_0, strides = x_87_strides_0, weight = layers_8_ffn_conv2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
            tensor<int32, [3]> x_89_perm_0 = const()[name = tensor<string, []>("x_89_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = x_87_cast_fp16)[name = tensor<string, []>("transpose_130")];
            tensor<fp16, [1, 110, 768]> input_183_cast_fp16 = add(x = input_175_cast_fp16, y = x_89_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<int32, [8]> k_padded_17_pad_0 = const()[name = tensor<string, []>("k_padded_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_17_mode_0 = const()[name = tensor<string, []>("k_padded_17_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_63_to_fp16 = const()[name = tensor<string, []>("const_63_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_17_cast_fp16 = pad(constant_val = const_63_to_fp16, mode = k_padded_17_mode_0, pad = k_padded_17_pad_0, x = k_33_cast_fp16)[name = tensor<string, []>("k_padded_17_cast_fp16")];
            tensor<int32, [8]> v_padded_17_pad_0 = const()[name = tensor<string, []>("v_padded_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_17_mode_0 = const()[name = tensor<string, []>("v_padded_17_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_64_to_fp16 = const()[name = tensor<string, []>("const_64_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_17_cast_fp16 = pad(constant_val = const_64_to_fp16, mode = v_padded_17_mode_0, pad = v_padded_17_pad_0, x = v_33_cast_fp16)[name = tensor<string, []>("v_padded_17_cast_fp16")];
            tensor<int32, []> var_1536_axis_0 = const()[name = tensor<string, []>("op_1536_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1536 = stack(axis = var_1536_axis_0, values = (k_padded_17_cast_fp16, v_padded_17_cast_fp16))[name = tensor<string, []>("op_1536_cast_fp16")];
            tensor<int32, []> var_1548 = const()[name = tensor<string, []>("op_1548"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_91_axes_0 = const()[name = tensor<string, []>("x_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_9_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134766272)))];
            tensor<fp16, []> var_1555_to_fp16 = const()[name = tensor<string, []>("op_1555_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_91_cast_fp16 = layer_norm(axes = x_91_axes_0, epsilon = var_1555_to_fp16, gamma = layers_9_norm_sa_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_9_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134767872)))];
            tensor<fp16, [1, 110, 2304]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_qkv_proj_weight_to_fp16, x = x_91_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<int32, [5]> var_1573 = const()[name = tensor<string, []>("op_1573"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_19_cast_fp16 = reshape(shape = var_1573, x = linear_45_cast_fp16)[name = tensor<string, []>("qkv_19_cast_fp16")];
            tensor<int32, [5]> q_55_begin_0 = const()[name = tensor<string, []>("q_55_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_55_end_0 = const()[name = tensor<string, []>("q_55_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_55_end_mask_0 = const()[name = tensor<string, []>("q_55_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_55_squeeze_mask_0 = const()[name = tensor<string, []>("q_55_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
            tensor<int32, [5]> k_37_begin_0 = const()[name = tensor<string, []>("k_37_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_37_end_0 = const()[name = tensor<string, []>("k_37_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_37_end_mask_0 = const()[name = tensor<string, []>("k_37_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_37_squeeze_mask_0 = const()[name = tensor<string, []>("k_37_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
            tensor<int32, [5]> v_37_begin_0 = const()[name = tensor<string, []>("v_37_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_37_end_0 = const()[name = tensor<string, []>("v_37_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_37_end_mask_0 = const()[name = tensor<string, []>("v_37_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_37_squeeze_mask_0 = const()[name = tensor<string, []>("v_37_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor<string, []>("v_37_cast_fp16")];
            tensor<int32, [4]> v_t_19_perm_0 = const()[name = tensor<string, []>("v_t_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1588_transpose_x_0 = const()[name = tensor<string, []>("op_1588_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1588_transpose_y_0 = const()[name = tensor<string, []>("op_1588_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = tensor<string, []>("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = tensor<string, []>("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_127")];
            tensor<fp16, [1, 12, 110, 64]> transpose_102 = transpose(perm = transpose_102_perm_0, x = q_55_cast_fp16)[name = tensor<string, []>("transpose_128")];
            tensor<fp16, [1, 12, 110, 110]> var_1588_cast_fp16 = matmul(transpose_x = var_1588_transpose_x_0, transpose_y = var_1588_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor<string, []>("op_1588_cast_fp16")];
            tensor<fp16, []> var_1589_to_fp16 = const()[name = tensor<string, []>("op_1589_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_73_cast_fp16 = mul(x = var_1588_cast_fp16, y = var_1589_to_fp16)[name = tensor<string, []>("attn_73_cast_fp16")];
            tensor<string, []> attn_73_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_73_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_73_cast_fp16_to_fp32 = cast(dtype = attn_73_cast_fp16_to_fp32_dtype_0, x = attn_73_cast_fp16)[name = tensor<string, []>("cast_75")];
            tensor<fp32, [1, 12, 110, 110]> input_185 = add(x = attn_73_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_185")];
            tensor<string, []> input_185_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_185_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_185_to_fp16 = cast(dtype = input_185_to_fp16_dtype_0, x = input_185)[name = tensor<string, []>("cast_74")];
            tensor<fp16, [1, 12, 110, 110]> attn_75_cast_fp16 = softmax(axis = var_1548, x = input_185_to_fp16)[name = tensor<string, []>("attn_75_cast_fp16")];
            tensor<bool, []> out_37_transpose_x_0 = const()[name = tensor<string, []>("out_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_37_transpose_y_0 = const()[name = tensor<string, []>("out_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_19_cast_fp16 = transpose(perm = v_t_19_perm_0, x = v_37_cast_fp16)[name = tensor<string, []>("transpose_129")];
            tensor<fp16, [1, 12, 110, 64]> out_37_cast_fp16 = matmul(transpose_x = out_37_transpose_x_0, transpose_y = out_37_transpose_y_0, x = attn_75_cast_fp16, y = v_t_19_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<int32, [4]> var_1600_perm_0 = const()[name = tensor<string, []>("op_1600_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1601 = const()[name = tensor<string, []>("op_1601"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_1600_cast_fp16 = transpose(perm = var_1600_perm_0, x = out_37_cast_fp16)[name = tensor<string, []>("transpose_126")];
            tensor<fp16, [1, 110, 768]> input_187_cast_fp16 = reshape(shape = var_1601, x = var_1600_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<fp16, [768, 768]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138306880)))];
            tensor<fp16, [1, 110, 768]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_46_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<int32, [1]> x_93_axes_0 = const()[name = tensor<string, []>("x_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_9_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139486592)))];
            tensor<fp16, [1, 110, 768]> x_93_cast_fp16 = layer_norm(axes = x_93_axes_0, epsilon = var_1555_to_fp16, gamma = layers_9_norm_xa_query_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("x_93_cast_fp16")];
            tensor<int32, [1]> memory_19_axes_0 = const()[name = tensor<string, []>("memory_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_9_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139488192)))];
            tensor<fp16, [1, 256, 768]> memory_19_cast_fp16 = layer_norm(axes = memory_19_axes_0, epsilon = var_1555_to_fp16, gamma = layers_9_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_19_cast_fp16")];
            tensor<fp16, [128, 768]> layers_9_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139489792)))];
            tensor<fp16, [1, 110, 128]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_9_cross_attn_q_proj_weight_to_fp16, x = x_93_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<int32, [4]> var_1624 = const()[name = tensor<string, []>("op_1624"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_1625_cast_fp16 = reshape(shape = var_1624, x = linear_47_cast_fp16)[name = tensor<string, []>("op_1625_cast_fp16")];
            tensor<fp16, [256, 768]> layers_9_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139686464)))];
            tensor<fp16, [1, 256, 256]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_9_cross_attn_kv_proj_weight_to_fp16, x = memory_19_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<int32, [5]> var_1629 = const()[name = tensor<string, []>("op_1629"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_19_cast_fp16 = reshape(shape = var_1629, x = linear_48_cast_fp16)[name = tensor<string, []>("kv_19_cast_fp16")];
            tensor<int32, [5]> var_1633_begin_0 = const()[name = tensor<string, []>("op_1633_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1633_end_0 = const()[name = tensor<string, []>("op_1633_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_1633_end_mask_0 = const()[name = tensor<string, []>("op_1633_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1633_squeeze_mask_0 = const()[name = tensor<string, []>("op_1633_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1633_cast_fp16 = slice_by_index(begin = var_1633_begin_0, end = var_1633_end_0, end_mask = var_1633_end_mask_0, squeeze_mask = var_1633_squeeze_mask_0, x = kv_19_cast_fp16)[name = tensor<string, []>("op_1633_cast_fp16")];
            tensor<int32, [5]> var_1637_begin_0 = const()[name = tensor<string, []>("op_1637_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_1637_end_0 = const()[name = tensor<string, []>("op_1637_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_1637_end_mask_0 = const()[name = tensor<string, []>("op_1637_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1637_squeeze_mask_0 = const()[name = tensor<string, []>("op_1637_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1637_cast_fp16 = slice_by_index(begin = var_1637_begin_0, end = var_1637_end_0, end_mask = var_1637_end_mask_0, squeeze_mask = var_1637_squeeze_mask_0, x = kv_19_cast_fp16)[name = tensor<string, []>("op_1637_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = tensor<string, []>("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1640_transpose_x_0 = const()[name = tensor<string, []>("op_1640_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1640_transpose_y_0 = const()[name = tensor<string, []>("op_1640_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_105 = transpose(perm = transpose_105_perm_0, x = var_1633_cast_fp16)[name = tensor<string, []>("transpose_123")];
            tensor<fp16, [1, 1, 110, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1625_cast_fp16)[name = tensor<string, []>("transpose_124")];
            tensor<fp16, [1, 1, 110, 256]> var_1640_cast_fp16 = matmul(transpose_x = var_1640_transpose_x_0, transpose_y = var_1640_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor<string, []>("op_1640_cast_fp16")];
            tensor<fp16, []> var_1641_to_fp16 = const()[name = tensor<string, []>("op_1641_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_77_cast_fp16 = mul(x = var_1640_cast_fp16, y = var_1641_to_fp16)[name = tensor<string, []>("attn_77_cast_fp16")];
            tensor<fp16, []> var_1539_to_fp16 = const()[name = tensor<string, []>("op_1539_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_191_cast_fp16 = select(a = var_1539_to_fp16, b = attn_77_cast_fp16, cond = var_151)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_79_cast_fp16 = softmax(axis = var_1548, x = input_191_cast_fp16)[name = tensor<string, []>("attn_79_cast_fp16")];
            tensor<bool, []> out_39_transpose_x_0 = const()[name = tensor<string, []>("out_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_39_transpose_y_0 = const()[name = tensor<string, []>("out_39_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_1637_cast_fp16)[name = tensor<string, []>("transpose_125")];
            tensor<fp16, [1, 1, 110, 128]> out_39_cast_fp16 = matmul(transpose_x = out_39_transpose_x_0, transpose_y = out_39_transpose_y_0, x = attn_79_cast_fp16, y = v_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<int32, [4]> var_1649_perm_0 = const()[name = tensor<string, []>("op_1649_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1650 = const()[name = tensor<string, []>("op_1650"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_1649_cast_fp16 = transpose(perm = var_1649_perm_0, x = out_39_cast_fp16)[name = tensor<string, []>("transpose_122")];
            tensor<fp16, [1, 110, 128]> input_193_cast_fp16 = reshape(shape = var_1650, x = var_1649_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<fp16, [768, 128]> layers_9_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140079744)))];
            tensor<fp16, [1, 110, 768]> linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_cross_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_49_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_9_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140276416)))];
            tensor<fp16, [1, 110, 768]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, epsilon = var_1555_to_fp16, gamma = layers_9_norm_ff_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
            tensor<int32, [3]> input_197_perm_0 = const()[name = tensor<string, []>("input_197_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_199_pad_type_0 = const()[name = tensor<string, []>("input_199_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_199_strides_0 = const()[name = tensor<string, []>("input_199_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_199_pad_0 = const()[name = tensor<string, []>("input_199_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_199_dilations_0 = const()[name = tensor<string, []>("input_199_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_199_groups_0 = const()[name = tensor<string, []>("input_199_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_9_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140278016)))];
            tensor<fp16, [1, 768, 110]> input_197_cast_fp16 = transpose(perm = input_197_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_121")];
            tensor<fp16, [1, 3072, 110]> input_199_cast_fp16 = conv(dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = layers_9_ffn_conv1_weight_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<string, []> input_201_mode_0 = const()[name = tensor<string, []>("input_201_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = input_199_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
            tensor<string, []> x_97_pad_type_0 = const()[name = tensor<string, []>("x_97_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_97_strides_0 = const()[name = tensor<string, []>("x_97_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_97_pad_0 = const()[name = tensor<string, []>("x_97_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_97_dilations_0 = const()[name = tensor<string, []>("x_97_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_97_groups_0 = const()[name = tensor<string, []>("x_97_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_9_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144996672)))];
            tensor<fp16, [1, 768, 110]> x_97_cast_fp16 = conv(dilations = x_97_dilations_0, groups = x_97_groups_0, pad = x_97_pad_0, pad_type = x_97_pad_type_0, strides = x_97_strides_0, weight = layers_9_ffn_conv2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
            tensor<int32, [3]> x_99_perm_0 = const()[name = tensor<string, []>("x_99_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_99_cast_fp16 = transpose(perm = x_99_perm_0, x = x_97_cast_fp16)[name = tensor<string, []>("transpose_120")];
            tensor<fp16, [1, 110, 768]> input_203_cast_fp16 = add(x = input_195_cast_fp16, y = x_99_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<int32, [8]> k_padded_19_pad_0 = const()[name = tensor<string, []>("k_padded_19_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_19_mode_0 = const()[name = tensor<string, []>("k_padded_19_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_70_to_fp16 = const()[name = tensor<string, []>("const_70_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_19_cast_fp16 = pad(constant_val = const_70_to_fp16, mode = k_padded_19_mode_0, pad = k_padded_19_pad_0, x = k_37_cast_fp16)[name = tensor<string, []>("k_padded_19_cast_fp16")];
            tensor<int32, [8]> v_padded_19_pad_0 = const()[name = tensor<string, []>("v_padded_19_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_19_mode_0 = const()[name = tensor<string, []>("v_padded_19_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_71_to_fp16 = const()[name = tensor<string, []>("const_71_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_19_cast_fp16 = pad(constant_val = const_71_to_fp16, mode = v_padded_19_mode_0, pad = v_padded_19_pad_0, x = v_37_cast_fp16)[name = tensor<string, []>("v_padded_19_cast_fp16")];
            tensor<int32, []> var_1702_axis_0 = const()[name = tensor<string, []>("op_1702_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1702 = stack(axis = var_1702_axis_0, values = (k_padded_19_cast_fp16, v_padded_19_cast_fp16))[name = tensor<string, []>("op_1702_cast_fp16")];
            tensor<int32, []> var_1714 = const()[name = tensor<string, []>("op_1714"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_10_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149715328)))];
            tensor<fp16, []> var_1721_to_fp16 = const()[name = tensor<string, []>("op_1721_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, epsilon = var_1721_to_fp16, gamma = layers_10_norm_sa_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_10_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149716928)))];
            tensor<fp16, [1, 110, 2304]> linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_qkv_proj_weight_to_fp16, x = x_101_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<int32, [5]> var_1739 = const()[name = tensor<string, []>("op_1739"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_21_cast_fp16 = reshape(shape = var_1739, x = linear_50_cast_fp16)[name = tensor<string, []>("qkv_21_cast_fp16")];
            tensor<int32, [5]> q_61_begin_0 = const()[name = tensor<string, []>("q_61_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_61_end_0 = const()[name = tensor<string, []>("q_61_end_0"), val = tensor<int32, [5]>([1, 110, 1, 12, 64])];
            tensor<bool, [5]> q_61_end_mask_0 = const()[name = tensor<string, []>("q_61_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_61_squeeze_mask_0 = const()[name = tensor<string, []>("q_61_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")];
            tensor<int32, [5]> k_41_begin_0 = const()[name = tensor<string, []>("k_41_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_41_end_0 = const()[name = tensor<string, []>("k_41_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_41_end_mask_0 = const()[name = tensor<string, []>("k_41_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_41_squeeze_mask_0 = const()[name = tensor<string, []>("k_41_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
            tensor<int32, [5]> v_41_begin_0 = const()[name = tensor<string, []>("v_41_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_41_end_0 = const()[name = tensor<string, []>("v_41_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_41_end_mask_0 = const()[name = tensor<string, []>("v_41_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_41_squeeze_mask_0 = const()[name = tensor<string, []>("v_41_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor<string, []>("v_41_cast_fp16")];
            tensor<int32, [4]> v_t_21_perm_0 = const()[name = tensor<string, []>("v_t_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1754_transpose_x_0 = const()[name = tensor<string, []>("op_1754_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1754_transpose_y_0 = const()[name = tensor<string, []>("op_1754_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = tensor<string, []>("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 110]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_117")];
            tensor<fp16, [1, 12, 110, 64]> transpose_106 = transpose(perm = transpose_106_perm_0, x = q_61_cast_fp16)[name = tensor<string, []>("transpose_118")];
            tensor<fp16, [1, 12, 110, 110]> var_1754_cast_fp16 = matmul(transpose_x = var_1754_transpose_x_0, transpose_y = var_1754_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor<string, []>("op_1754_cast_fp16")];
            tensor<fp16, []> var_1755_to_fp16 = const()[name = tensor<string, []>("op_1755_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 110, 110]> attn_81_cast_fp16 = mul(x = var_1754_cast_fp16, y = var_1755_to_fp16)[name = tensor<string, []>("attn_81_cast_fp16")];
            tensor<string, []> attn_81_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("attn_81_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 110, 110]> attn_81_cast_fp16_to_fp32 = cast(dtype = attn_81_cast_fp16_to_fp32_dtype_0, x = attn_81_cast_fp16)[name = tensor<string, []>("cast_73")];
            tensor<fp32, [1, 12, 110, 110]> input_205 = add(x = attn_81_cast_fp16_to_fp32, y = var_102)[name = tensor<string, []>("input_205")];
            tensor<string, []> input_205_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_205_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 110, 110]> input_205_to_fp16 = cast(dtype = input_205_to_fp16_dtype_0, x = input_205)[name = tensor<string, []>("cast_72")];
            tensor<fp16, [1, 12, 110, 110]> attn_83_cast_fp16 = softmax(axis = var_1714, x = input_205_to_fp16)[name = tensor<string, []>("attn_83_cast_fp16")];
            tensor<bool, []> out_41_transpose_x_0 = const()[name = tensor<string, []>("out_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_41_transpose_y_0 = const()[name = tensor<string, []>("out_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 110, 64]> v_t_21_cast_fp16 = transpose(perm = v_t_21_perm_0, x = v_41_cast_fp16)[name = tensor<string, []>("transpose_119")];
            tensor<fp16, [1, 12, 110, 64]> out_41_cast_fp16 = matmul(transpose_x = out_41_transpose_x_0, transpose_y = out_41_transpose_y_0, x = attn_83_cast_fp16, y = v_t_21_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<int32, [4]> var_1766_perm_0 = const()[name = tensor<string, []>("op_1766_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1767 = const()[name = tensor<string, []>("op_1767"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 12, 64]> var_1766_cast_fp16 = transpose(perm = var_1766_perm_0, x = out_41_cast_fp16)[name = tensor<string, []>("transpose_116")];
            tensor<fp16, [1, 110, 768]> input_207_cast_fp16 = reshape(shape = var_1767, x = var_1766_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<fp16, [768, 768]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153255936)))];
            tensor<fp16, [1, 110, 768]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_209_cast_fp16 = add(x = input_203_cast_fp16, y = linear_51_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<int32, [1]> x_103_axes_0 = const()[name = tensor<string, []>("x_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_10_norm_xa_query_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_norm_xa_query_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154435648)))];
            tensor<fp16, [1, 110, 768]> x_103_cast_fp16 = layer_norm(axes = x_103_axes_0, epsilon = var_1721_to_fp16, gamma = layers_10_norm_xa_query_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")];
            tensor<int32, [1]> memory_21_axes_0 = const()[name = tensor<string, []>("memory_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_10_norm_xa_memory_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_norm_xa_memory_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154437248)))];
            tensor<fp16, [1, 256, 768]> memory_21_cast_fp16 = layer_norm(axes = memory_21_axes_0, epsilon = var_1721_to_fp16, gamma = layers_10_norm_xa_memory_weight_to_fp16, x = encoder_output)[name = tensor<string, []>("memory_21_cast_fp16")];
            tensor<fp16, [128, 768]> layers_10_cross_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_cross_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [128, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154438848)))];
            tensor<fp16, [1, 110, 128]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_10_cross_attn_q_proj_weight_to_fp16, x = x_103_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<int32, [4]> var_1790 = const()[name = tensor<string, []>("op_1790"), val = tensor<int32, [4]>([1, 110, 1, 128])];
            tensor<fp16, [1, 110, 1, 128]> var_1791_cast_fp16 = reshape(shape = var_1790, x = linear_52_cast_fp16)[name = tensor<string, []>("op_1791_cast_fp16")];
            tensor<fp16, [256, 768]> layers_10_cross_attn_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_cross_attn_kv_proj_weight_to_fp16"), val = tensor<fp16, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154635520)))];
            tensor<fp16, [1, 256, 256]> linear_53_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = layers_10_cross_attn_kv_proj_weight_to_fp16, x = memory_21_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<int32, [5]> var_1795 = const()[name = tensor<string, []>("op_1795"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<fp16, [1, 256, 2, 1, 128]> kv_21_cast_fp16 = reshape(shape = var_1795, x = linear_53_cast_fp16)[name = tensor<string, []>("kv_21_cast_fp16")];
            tensor<int32, [5]> var_1799_begin_0 = const()[name = tensor<string, []>("op_1799_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1799_end_0 = const()[name = tensor<string, []>("op_1799_end_0"), val = tensor<int32, [5]>([1, 256, 1, 1, 128])];
            tensor<bool, [5]> var_1799_end_mask_0 = const()[name = tensor<string, []>("op_1799_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1799_squeeze_mask_0 = const()[name = tensor<string, []>("op_1799_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1799_cast_fp16 = slice_by_index(begin = var_1799_begin_0, end = var_1799_end_0, end_mask = var_1799_end_mask_0, squeeze_mask = var_1799_squeeze_mask_0, x = kv_21_cast_fp16)[name = tensor<string, []>("op_1799_cast_fp16")];
            tensor<int32, [5]> var_1803_begin_0 = const()[name = tensor<string, []>("op_1803_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> var_1803_end_0 = const()[name = tensor<string, []>("op_1803_end_0"), val = tensor<int32, [5]>([1, 256, 2, 1, 128])];
            tensor<bool, [5]> var_1803_end_mask_0 = const()[name = tensor<string, []>("op_1803_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> var_1803_squeeze_mask_0 = const()[name = tensor<string, []>("op_1803_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 256, 1, 128]> var_1803_cast_fp16 = slice_by_index(begin = var_1803_begin_0, end = var_1803_end_0, end_mask = var_1803_end_mask_0, squeeze_mask = var_1803_squeeze_mask_0, x = kv_21_cast_fp16)[name = tensor<string, []>("op_1803_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = tensor<string, []>("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1806_transpose_x_0 = const()[name = tensor<string, []>("op_1806_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1806_transpose_y_0 = const()[name = tensor<string, []>("op_1806_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = tensor<string, []>("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, 128, 256]> transpose_109 = transpose(perm = transpose_109_perm_0, x = var_1799_cast_fp16)[name = tensor<string, []>("transpose_113")];
            tensor<fp16, [1, 1, 110, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1791_cast_fp16)[name = tensor<string, []>("transpose_114")];
            tensor<fp16, [1, 1, 110, 256]> var_1806_cast_fp16 = matmul(transpose_x = var_1806_transpose_x_0, transpose_y = var_1806_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor<string, []>("op_1806_cast_fp16")];
            tensor<fp16, []> var_1807_to_fp16 = const()[name = tensor<string, []>("op_1807_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 110, 256]> attn_85_cast_fp16 = mul(x = var_1806_cast_fp16, y = var_1807_to_fp16)[name = tensor<string, []>("attn_85_cast_fp16")];
            tensor<fp16, []> var_1705_to_fp16 = const()[name = tensor<string, []>("op_1705_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, [1, 1, 110, 256]> input_211_cast_fp16 = select(a = var_1705_to_fp16, b = attn_85_cast_fp16, cond = var_151)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<fp16, [1, 1, 110, 256]> attn_87_cast_fp16 = softmax(axis = var_1714, x = input_211_cast_fp16)[name = tensor<string, []>("attn_87_cast_fp16")];
            tensor<bool, []> out_43_transpose_x_0 = const()[name = tensor<string, []>("out_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_43_transpose_y_0 = const()[name = tensor<string, []>("out_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 256, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_1803_cast_fp16)[name = tensor<string, []>("transpose_115")];
            tensor<fp16, [1, 1, 110, 128]> out_43_cast_fp16 = matmul(transpose_x = out_43_transpose_x_0, transpose_y = out_43_transpose_y_0, x = attn_87_cast_fp16, y = v_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<int32, [4]> var_1815_perm_0 = const()[name = tensor<string, []>("op_1815_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1816 = const()[name = tensor<string, []>("op_1816"), val = tensor<int32, [3]>([1, 110, -1])];
            tensor<fp16, [1, 110, 1, 128]> var_1815_cast_fp16 = transpose(perm = var_1815_perm_0, x = out_43_cast_fp16)[name = tensor<string, []>("transpose_112")];
            tensor<fp16, [1, 110, 128]> input_213_cast_fp16 = reshape(shape = var_1816, x = var_1815_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<fp16, [768, 128]> layers_10_cross_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_cross_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [768, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155028800)))];
            tensor<fp16, [1, 110, 768]> linear_54_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_cross_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, [1, 110, 768]> input_215_cast_fp16 = add(x = input_209_cast_fp16, y = linear_54_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<int32, [1]> x_105_axes_0 = const()[name = tensor<string, []>("x_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_10_norm_ff_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_norm_ff_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155225472)))];
            tensor<fp16, [1, 110, 768]> x_105_cast_fp16 = layer_norm(axes = x_105_axes_0, epsilon = var_1721_to_fp16, gamma = layers_10_norm_ff_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_105_cast_fp16")];
            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_219_pad_type_0 = const()[name = tensor<string, []>("input_219_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> input_219_strides_0 = const()[name = tensor<string, []>("input_219_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_219_pad_0 = const()[name = tensor<string, []>("input_219_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_219_dilations_0 = const()[name = tensor<string, []>("input_219_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_219_groups_0 = const()[name = tensor<string, []>("input_219_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [3072, 768, 1]> layers_10_ffn_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_ffn_conv1_weight_to_fp16"), val = tensor<fp16, [3072, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155227072)))];
            tensor<fp16, [1, 768, 110]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_105_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 3072, 110]> input_219_cast_fp16 = conv(dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = layers_10_ffn_conv1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<string, []> input_221_mode_0 = const()[name = tensor<string, []>("input_221_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 3072, 110]> input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<string, []> x_107_pad_type_0 = const()[name = tensor<string, []>("x_107_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_107_strides_0 = const()[name = tensor<string, []>("x_107_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_107_pad_0 = const()[name = tensor<string, []>("x_107_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_107_dilations_0 = const()[name = tensor<string, []>("x_107_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_107_groups_0 = const()[name = tensor<string, []>("x_107_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3072, 1]> layers_10_ffn_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_ffn_conv2_weight_to_fp16"), val = tensor<fp16, [768, 3072, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159945728)))];
            tensor<fp16, [1, 768, 110]> x_107_cast_fp16 = conv(dilations = x_107_dilations_0, groups = x_107_groups_0, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = x_107_strides_0, weight = layers_10_ffn_conv2_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
            tensor<int32, [3]> x_109_perm_0 = const()[name = tensor<string, []>("x_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 110, 768]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = x_107_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [1, 110, 768]> input_223_cast_fp16 = add(x = input_215_cast_fp16, y = x_109_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
            tensor<int32, [8]> k_padded_21_pad_0 = const()[name = tensor<string, []>("k_padded_21_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_21_mode_0 = const()[name = tensor<string, []>("k_padded_21_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_77_to_fp16 = const()[name = tensor<string, []>("const_77_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_21_cast_fp16 = pad(constant_val = const_77_to_fp16, mode = k_padded_21_mode_0, pad = k_padded_21_pad_0, x = k_41_cast_fp16)[name = tensor<string, []>("k_padded_21_cast_fp16")];
            tensor<int32, [8]> v_padded_21_pad_0 = const()[name = tensor<string, []>("v_padded_21_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_21_mode_0 = const()[name = tensor<string, []>("v_padded_21_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_78_to_fp16 = const()[name = tensor<string, []>("const_78_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_21_cast_fp16 = pad(constant_val = const_78_to_fp16, mode = v_padded_21_mode_0, pad = v_padded_21_pad_0, x = v_41_cast_fp16)[name = tensor<string, []>("v_padded_21_cast_fp16")];
            tensor<int32, []> var_1868_axis_0 = const()[name = tensor<string, []>("op_1868_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1868 = stack(axis = var_1868_axis_0, values = (k_padded_21_cast_fp16, v_padded_21_cast_fp16))[name = tensor<string, []>("op_1868_cast_fp16")];
            tensor<int32, [1]> x_111_axes_0 = const()[name = tensor<string, []>("x_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> layers_11_norm_sa_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_norm_sa_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164664384)))];
            tensor<fp16, []> var_1884_to_fp16 = const()[name = tensor<string, []>("op_1884_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 110, 768]> x_111_cast_fp16 = layer_norm(axes = x_111_axes_0, epsilon = var_1884_to_fp16, gamma = layers_11_norm_sa_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
            tensor<fp16, [2304, 768]> layers_11_self_attn_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_qkv_proj_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164665984)))];
            tensor<fp16, [1, 110, 2304]> linear_55_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_qkv_proj_weight_to_fp16, x = x_111_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<int32, [5]> var_1897 = const()[name = tensor<string, []>("op_1897"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<fp16, [1, 110, 3, 12, 64]> qkv_cast_fp16 = reshape(shape = var_1897, x = linear_55_cast_fp16)[name = tensor<string, []>("qkv_cast_fp16")];
            tensor<int32, [5]> k_45_begin_0 = const()[name = tensor<string, []>("k_45_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_45_end_0 = const()[name = tensor<string, []>("k_45_end_0"), val = tensor<int32, [5]>([1, 110, 2, 12, 64])];
            tensor<bool, [5]> k_45_end_mask_0 = const()[name = tensor<string, []>("k_45_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_45_squeeze_mask_0 = const()[name = tensor<string, []>("k_45_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
            tensor<int32, [5]> v_45_begin_0 = const()[name = tensor<string, []>("v_45_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_45_end_0 = const()[name = tensor<string, []>("v_45_end_0"), val = tensor<int32, [5]>([1, 110, 3, 12, 64])];
            tensor<bool, [5]> v_45_end_mask_0 = const()[name = tensor<string, []>("v_45_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_45_squeeze_mask_0 = const()[name = tensor<string, []>("v_45_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp16, [1, 110, 12, 64]> v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor<string, []>("v_45_cast_fp16")];
            tensor<int32, [8]> k_padded_pad_0 = const()[name = tensor<string, []>("k_padded_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> k_padded_mode_0 = const()[name = tensor<string, []>("k_padded_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_81_to_fp16 = const()[name = tensor<string, []>("const_81_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> k_padded_cast_fp16 = pad(constant_val = const_81_to_fp16, mode = k_padded_mode_0, pad = k_padded_pad_0, x = k_45_cast_fp16)[name = tensor<string, []>("k_padded_cast_fp16")];
            tensor<int32, [8]> v_padded_pad_0 = const()[name = tensor<string, []>("v_padded_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 402, 0, 0, 0, 0])];
            tensor<string, []> v_padded_mode_0 = const()[name = tensor<string, []>("v_padded_mode_0"), val = tensor<string, []>("constant")];
            tensor<fp16, []> const_82_to_fp16 = const()[name = tensor<string, []>("const_82_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 512, 12, 64]> v_padded_cast_fp16 = pad(constant_val = const_82_to_fp16, mode = v_padded_mode_0, pad = v_padded_pad_0, x = v_45_cast_fp16)[name = tensor<string, []>("v_padded_cast_fp16")];
            tensor<int32, []> var_1958_axis_0 = const()[name = tensor<string, []>("op_1958_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [2, 1, 512, 12, 64]> var_1958 = stack(axis = var_1958_axis_0, values = (k_padded_cast_fp16, v_padded_cast_fp16))[name = tensor<string, []>("op_1958_cast_fp16")];
        } -> (var_208, var_374, var_540, var_706, var_872, var_1038, var_1204, var_1370, var_1536, var_1702, var_1868, var_1958);
}